LinkGuard cover — On-page SEO optimization checklist: 35 items for the 2026 audit

The 2026 on-page SEO audit is 35 items across 7 categories. The single highest-leverage fix — drops Google's title-rewrite rate on your page from ~76% to ~20.6% in one change — is making your title tag and your H1 thematically coherent. Source: John McAlpin / Search Engine Land Q1 2025 study, ~30,000 keywords.

The page you spent two weeks on ranked sixth for a week, dropped to nineteenth after a core update, and the Slack message from your CEO says "what happened to that piece." You open the URL, look at the title, look at the H1, look at the lead paragraph. The title is sharp. The H1 is sharp. The lead paragraph spends 220 words setting up before answering anything. The AI Overview Google built on top of the SERP cites three competitors and not you. Your meta description, which you spent twenty minutes crafting, is rewritten in the SERP to something Google pulled out of the page body.

This is the 2026 on-page audit you're about to run. It is not 2018. The fundamentals are still mostly intact — unique titles, unique H1s, descriptive metas, clean heading hierarchy, alt text on images, internal links with descriptive anchors, schema where Google supports rich results. What has shifted in the last 18 months is sharp enough to make a 2022 audit guide actively misleading on five fronts: Google rewrites the majority of titles anyway, the Helpful Content Update was folded into core ranking in March 2024 (there is no separate sitewide classifier to wait for any more), AI Overviews and ChatGPT and Perplexity chunk your page into passages and retrieve passages, schema markup is being actively retired (FAQ rich results died May 2026, HowTo died in 2023, seven more types in June 2025), and "longer content always ranks better" was correlation that the post-HCU Navboost world has actively punished.

This checklist is what to ship in 2026 and what to retire. Your work over the next read-through is the difference between a page that ranks AND gets cited AI-side, and a page that does one or the other. The on-page win is no longer a single signal — it's six signals layered (title-H1 coherence, post-HCU content quality, AI-extractable structure, schema discipline, image-text-image fusion, internal-link anchor variety) and the page either passes all six or rakes in a fraction of what it could.

Three numbers anchor what's at stake.

1. Google rewrote 76.04% of all title tags in Q1 2025 — only ~24% of titles survived unchanged. When title and H1 are coherent (thematically aligned, not necessarily identical), the rewrite rate drops to ~20.6%. Source: John McAlpin study of ~30,000 keywords across YMYL and non-YMYL, top-20 results per query, published in Search Engine Land May 1, 2025. Brand names get removed in 63% of modified titles; Google retains only ~35% of the original content on average. Translation for your audit: stop micro-engineering 60-character clickbait titles. Write a clear title that matches the H1 thematically, and Google leaves you alone the most often. The single highest-leverage on-page lever in 2026 is title-H1 coherence.

2. The Helpful Content Update was deprecated and absorbed into core ranking in March 2024 — there is no separate "helpful" sitewide classifier any more. Helpfulness is evaluated continuously, at page AND passage level, on every query. Source: Lily Ray / Amsive analysis of the absorption. Operator consequences: (a) recovery is continuous, not periodic — you cannot "wait for the next HCU run"; (b) page-level on-page work matters more again because passage-level signals are now evaluated per chunk; (c) only ~22% of HCU-hit sites tracked by Glenn Gabe through August 2024 showed a 20%+ traffic lift post the August 2024 core update — recovery is real but slow and conditional on shipping the post-HCU quality changes.

3. ~55% of AI Overview citations come from the top 30% of the page; ~24% from the middle 30-60%; ~21% from the bottom. Source: Search Engine Land "answer-first content" guide. Cyrus Shepard / Zyppy 2026 winners study adds the structural lever: the top-correlated winning trait across 400+ sites is "offering a product or service" (0.391), then "allowing task completion" (0.381), then "owning proprietary assets" (0.357) — functional differentiation, not informational depth. Translation: front-load the answer, ship answer-first paragraphs in the top 30% of the page, and prove the page does something — not just describes something.

This checklist has 35 items across 7 categories: title + meta + H1 fundamentals, post-HCU content quality, structure for AI extraction, structured data (ship what's supported), images and media, internal linking patterns, and anti-patterns being sold as on-page best practice in 2026 that don't survive scrutiny. Tier filter, browser-saved progress, no account required.

Vocabulary

If you have read our GEO readiness checklist, technical SEO pre-launch checklist, and site migration checklist you have most of these. New terms in italics.

Title rewrite rate: the share of your titles Google replaces in the SERP with something else (its own generation from your H1 / page body / anchor text). 76% in Q1 2025. Drops to ~20% when title and H1 align.
Title-H1 coherence: the property that your title tag and your H1 make the same promise about the page. They do not have to be identical — they have to agree thematically. The single biggest 2026 lever to keep Google from rewriting your title.
HCU (Helpful Content Update): Google's 2022-2023 sitewide quality classifier. Deprecated March 2024 — folded into core ranking. There is no separate HCU signal any more.
Navboost: Google's user-behavior-signal system that uses ~13 months of click and dwell data per query. Confirmed in the 2024 antitrust trial documents. Post-HCU absorption, Navboost is the proximate reason "bouncing because the answer is buried" hurts ranking.
Passage retrieval / chunk-level relevance: AI search engines (ChatGPT Search, Claude, Perplexity, Google AI Overviews) do not retrieve whole pages — they break pages into passages and retrieve the most relevant passages for synthesis. Per Aleyda Solís, this is THE structural change every on-page audit must serve in 2026.
Answer-first paragraph: a paragraph that answers the section's primary question in the first 1-2 sentences (30-50 words) before expanding. Replaces the traditional "build up to the answer" essay style.
Entity salience: Google's NLP score for how prominently a topic / entity is featured on a page. 0.0-1.0 scale; >0.10 = present, >0.30 = strong topical focus. Replaces "keyword density" as the operator metric for on-page topical coverage.
Scaled content abuse: Google's policy term for mass-produced content with low originality / no human gate. Manual-action risk live since June 2025. AI assistance to a human editor is fine; LLM publishing at scale with no human gate is the policy violation.
Schema rich result vs schema as data: a rich result is a SERP feature (FAQ dropdown, HowTo step list, review stars). Schema as data is the structured metadata. Many rich-result features have been retired in 2025-2026 (FAQ, HowTo, +7 more) while the underlying schema types remain valid — operators should ship schema where Google's Rich Results docs list it as supported, not where Schema.org defines it.

Quick answers (the 4 questions you searched for)

Why does Google rewrite my title tags? Google rewrote 76.04% of all title tags in Q1 2025 per John McAlpin's ~30,000-keyword study published in Search Engine Land. Brand names get removed in 63% of modified titles. The single biggest lever to stop the rewrite: make your title tag and your H1 thematically coherent — when they align, the rewrite rate drops from ~76% to ~20.6%. Write a clear title that matches your H1, and Google leaves you alone the most often.

Is keyword density still a ranking factor in 2026? No. John Mueller confirmed keyword density is not a ranking factor as early as 2011 and has restated it multiple times since. Modern Google uses NLP entity-salience scoring (0.0-1.0 scale; >0.10 = present, >0.30 = strong topical focus). The honest operator metric for on-page topical coverage is entity-coverage / topical-completeness — does your page mention the entities a knowledgeable reader would expect on this topic — not density percentage.

Should I still ship FAQ schema if rich results are deprecated? Optional. Google deprecated FAQ rich results on May 7, 2026 — the SERP feature is gone. Google has been clear that unused schema doesn't hurt and FAQPage remains a valid Schema.org type. If you have real Q&A content on the page (genuine user questions, genuine answers), marking it up is fine. Don't manufacture FAQ blocks chasing the (gone) SERP feature OR chasing the unverified vendor claim that FAQ schema gives 2.8-3.2x AI Overview citation lift — that claim has no published methodology behind it.

Will Google penalize AI-generated content? Not as a category — but yes for "scaled content abuse." Google's policy distinguishes AI assistance to a human editor (fine) from mass-produced content with low originality and no human gate (manual-action risk since June 2025). Glenn Gabe, December 2025 core-update analysis: "publishing AI content without human intervention is extremely risky; humans should be involved in the process (editing, reviewing, fact-checking)." The Navboost user-signal layer adds a separate risk — scaled thin content that gets bounced is an active negative regardless of whether the content was AI- or human-generated.

What success looks like

+90 days post-audit: your priority pages either rank higher than they did pre-audit OR get cited in at least one AI surface (Google AI Overview / AI Mode, ChatGPT Search, Perplexity) for at least one priority query. Your title-rewrite rate on the audited URLs in GSC's Performance report shows the original titles being kept in >75% of impressions (vs the 24% baseline). Your internal-anchor portfolio is varied per the Zyppy 23M-link study pattern, not the same exact-match phrase repeated. Your schema is the supported set only, validated in Google's Rich Results Test. The two MOST important things you do NOT do: pad pages to hit a word count, and ship AI-generated content without a human gate. Beating an expectation you set is a career-defining moment; missing one you didn't set is a career-defining incident.

0 / 35 · 0%

Critical Important Nice to have

Show only undone

The parts Google is most likely to rewrite (titles 76% of the time in Q1 2025, meta ~70%). The 2026 game is not 'perfect title' — it's 'close enough to H1 that Google leaves it alone.'

Critical Title and H1 are thematically coherent — same promise, not necessarily identical wording John McAlpin's Q1 2025 study (~30,000 keywords, top-20 results per query, via Advanced Web Ranking + Screaming Frog) found Google rewrote 76.04% of title tags overall — but only 20.6% of titles where the title and H1 were thematically coherent. That's the single highest-leverage 2026 on-page lever. The 'coherent' bar is low: same primary topic phrase, same intent. They do not have to be identical.
How to do this
Audit:
Run a Screaming Frog crawl of priority URLs, export title + H1 as adjacent columns.
For each URL: do they make the SAME promise about the page? Same primary topic phrase? Same intent (informational / commercial / navigational)?
If not: rewrite the title to match the H1's thematic frame (not its literal wording). Example bad: title 'Best CRM tools for SaaS startups in 2026' / H1 'How to pick a CRM in 2026'. Fix: title 'How to pick a CRM in 2026 (SaaS-startup perspective)'.
Re-measure rewrite rate at +30 days: GSC doesn't expose the rewritten SERP title directly. Sample 20 priority URLs in an incognito SERP for each page's target query, copy the displayed title vs your declared title into a sheet, calculate the keep rate. Coherent pages should keep the original >75% of the time.
Run the on-page audit free
Important Title tag length: 51-60 characters (the rewrite-resistance sweet spot) McAlpin's data: titles in the 51-60 character window have the lowest rewrite rate. Below 30 chars Google often appends the brand. Above 60 chars Google truncates and frequently rewrites to its own preferred length. 51-60 chars is enough to express the topic + intent without triggering the truncation-rewrite cascade.
How to do this
Audit:
Crawl all priority URLs, sort by title length.
Flag titles < 30 chars or > 60 chars for rewrite.
Target 51-60 chars (~55 is the operator-safe midpoint).
Counting trick: every character counts including spaces. Pipes |, em-dashes —, and middle dots · are each ~1.3 visual chars in Google's SERP — use sparingly.
Don't fight the limit by removing the brand if your audience expects to see it. Google removes the brand in 63% of rewrites anyway, so for a high-trust YMYL page it's defensible to keep brand at the END (where Google's removal is least disruptive to meaning).
Run the on-page audit free
Important Single H1 per page, containing the page's primary topic phrase Mueller, repeatedly: 'You can use H1 tags as often as you want on a page... your site is going to rank perfectly fine with no H1 or with five H1s.' Multiple H1s do not hurt rankings. BUT: 71.6% of screen-reader users navigate webpages by jumping between headings (WebAIM). One H1 is the HTML5 spec + accessibility baseline. And it's the cleanest signal for AI passage retrieval — chunked extraction works better when the page has one unambiguous H1 topic.
How to do this
Ship one H1 per page. The H1 should contain the page's primary topic phrase naturally (not stuffed). Example: 'How to pick a CRM in 2026' (good) vs 'Best Best CRM Software Tools CRM CRM 2026' (stuffed; over-optimization signal).
What to check:
Exactly one <h1> per page (use Screaming Frog → H1 → 'Multiple' filter).
H1 not empty.
H1 not duplicated across many pages (every page should have a unique H1).
H1 contains the primary topic phrase the page targets.
If the H1 is empty or missing on the homepage / landing page specifically: that's a frequent CMS theme bug worth flagging.
Run the on-page audit free
Important Meta description ≤155 chars, written for the cases Google keeps it (branded + specific informational) Google rewrites meta descriptions ~62-87% of the time depending on study (Ahrefs 2020: 62.78%; Portent 2024: 87%; recent tracking ~60-70%). It's not a ranking factor (Google has been clear since 2009) and it's not the 2017 CTR lever it was. But Google keeps your description on branded queries and high-intent informational queries where the user's intent is closest to your framing — and that's roughly a third of impressions on a high-traffic page. Ship it; write it for those cases.
How to do this
Audit:
Every indexable page has a unique meta description.
Length ≤ 155 characters (Google truncates around 155-160 on desktop; mobile truncates earlier).
Front-loads the value proposition / answer (since truncation cuts the tail).
Includes the primary topic phrase naturally — for branded / specific-informational queries where Google keeps your description, this matters.
What NOT to do:
Don't stuff keywords (Google may reject the description entirely and write its own).
Don't write generic boilerplate that could apply to any page on your site (Google will rewrite from page content anyway).
Don't ship missing metas on important pages — leaving Google to generate from page text gives you zero control over the ~30% of impressions where the original WOULD be kept.
Run the on-page audit free
Nice URL slug short, descriptive, kebab-case — matches the H1 conceptually URL slugs are a weak ranking signal but a real UX signal. Short descriptive slugs get shared and clicked more in chat / forum / email contexts because the URL itself communicates what the page is. Slug should match the H1 conceptually (not literally), use kebab-case (no underscores, no random hashes), and be short enough to share verbally without 'and then there's a bunch of letters' framing.
How to do this
Slug rules:
Kebab-case (on-page-seo-checklist), not snake_case, not camelCase, not random hashes.
Short (3-6 words usually).
Matches H1 conceptually — same primary phrase or its sluggified equivalent.
No stop words unless they materially change meaning (guide, tips, how-to are usually droppable).
Stable: do not change slugs of indexed pages without a 301 redirect to the new slug.
Bad examples: /p/1234567, /best_on_page_SEO_Guide_2026, /articles/2026/05/some-very-long-slug-that-includes-every-target-keyword-you-could-think-of.
Good example: /blog/on-page-seo-optimization-checklist.
Critical Self-referencing canonical present and pointing at the page's intended canonical URL Missing or cross-domain canonicals nuke pages out of the index — the #1 most common bug surfaced in real on-page audits per operator consensus. Google honours the canonical signal; if it says 'this page is canonical elsewhere,' Google de-indexes this URL. Self-referencing canonical is the operator default and is the cheapest insurance against accidental URL-variant duplication (tracking params, A/B variants, trailing-slash drift).
How to do this
Audit:
Crawl with Screaming Frog → Canonicals tab.
For every indexable URL the canonical MUST: (a) be present in raw HTML (not JS-injected), (b) be absolute (https://example.com/page, not /page), (c) match the page URL itself (self-referencing).
Spot-check 20 priority pages: curl -s https://example.com/page | grep canonical.
Common bugs:
Canonical points at the OLD domain after a migration (see migration checklist idx-02).
Canonical points at the homepage from every page (template bug).
JS-injected canonical that Googlebot may not process consistently — must be in raw HTML.
Validate this canonical free
Critical No accidental noindex on pages you want indexed — meta tag, X-Robots-Tag header, OR CMS plugin defaults Three independent surfaces can carry noindex (HTML meta tag, HTTP X-Robots-Tag header, per-CMS plugin defaults). Any one shipped from staging silently de-indexes. The #1 cause of 'why did this page disappear after a deploy' tickets — Google honours the strictest signal, so the directive only has to leak in one place to cost you the page entirely.
How to do this
Three checks per priority URL:
Meta tag: view-source the page; search for noindex. Should be absent (except on intentionally-noindex'd surfaces like cart, account, search-results, internal admin).
HTTP header: curl -sI https://example.com/page | grep -i x-robots-tag — should return nothing or no noindex value.
GSC URL Inspection: 'Indexing allowed: yes' on every indexable URL.
WordPress site? Check Settings → Reading → 'Discourage search engines from indexing this site' — must be UNCHECKED on production. This is the most common WP-specific noindex leak in 2026 deploy disasters.
Cross-link our indexation troubleshooting checklist for the full diagnostic when a page mysteriously vanishes from the index.
Run the on-page audit free

HCU was absorbed into core ranking March 2024 — there is no separate sitewide classifier to wait for. Helpfulness is evaluated continuously at page AND passage level. The HCU-recovery case studies all point at the same on-page levers.

Critical Author byline + credentials + Person/ProfilePage schema with sameAs cluster Marie Haynes's February 2026 analysis of sites that improved post-December 2025 core update: trust signals were the #1 differentiator — reliable customer service, author credentials, positive reputation. Aleyda Solís's July 2025 AI Search checklist lists structured author data as non-negotiable. Anonymous 'team' bylines are an E-E-A-T deficit. A named author with bio + credentials + sameAs links to LinkedIn / professional bodies is the cheapest E-E-A-T win available.
How to do this
For every article:
Visible author byline at the top — full name, role, credentials in one line.
Author bio page exists at a stable URL (e.g. /authors/firstname-lastname/) with Person + ProfilePage schema.
The Person schema has sameAs array pointing to LinkedIn, professional bodies, X / academic profiles where applicable.
Article's JSON-LD author property references the author URL, not just the name.
Validation: paste the article URL into Rich Results Test. The Article block should resolve to a Person object with non-empty sameAs. See sibling pre-launch checklist sd-02 for the schema pattern.
Run the on-page audit free
Critical Content reflects first-hand experience — original photos, original test data, real examples Marie Haynes's post-Dec-2025 analysis of sites that improved: original content beat commodity information — firsthand reviews, unique insights, original photos and videos. Cyrus Shepard's 2026 winners study (400+ sites) found the top winning traits are 'offering a product or service' (0.391 correlation), 'allowing task completion' (0.381), and 'owning proprietary assets' (0.357). Translation: pages that show first-hand experience win; pages that paraphrase what's already on the web lose.
How to do this
For each priority page, audit honestly:
Are the screenshots original or stock?
Are the test results / data points your own, or quoted from someone else?
Is there at least one concrete example from your own experience?
Could the same page have been written by someone who has never used the thing?
If the answer is 'mostly paraphrase / stock': budget 2-4 hours per priority page to inject genuine first-hand material. One original screenshot, one real test result, one concrete experience-anchor is enough to clear the bar most paraphrase-pages don't.
This is the single biggest 2026 content-quality lever — and the cheapest. The competition can't fake first-hand experience.
Critical No scaled AI content shipped without a human editor gate Google's policy distinguishes AI assistance (fine) from scaled content abuse (manual-action risk live since June 2025). Glenn Gabe, December 2025 core-update analysis: 'publishing AI content without human intervention is extremely risky; humans should be involved (editing, reviewing, fact-checking).' Separate Navboost risk: scaled thin content gets bounced, which is an active ranking negative regardless of whether the text was AI- or human-written.
How to do this
Policy for your content team:
AI assistance to a human editor: fine. Drafts, outlines, transcript cleanup, idea-expansion — all OK.
LLM-to-publish with no human gate: never. Manual-action risk + Navboost risk + scaled-content-abuse policy risk all at once.
If you ARE shipping AI-assisted content: log human editor time per piece. If the average human-edit time is under ~20 minutes per 1,500 words, you're in the 'scaled' policy zone whether the editor signed off or not.
Recovery from a scaled-content-abuse manual action takes months, requires deleting the offending pages, and (per Glenn Gabe's recovery framing) only works if Google sees the cleanup sustained over time. Don't risk it for the cost saving.
Important Topic completeness — entity coverage matches what a knowledgeable reader expects Modern Google uses NLP entity-salience scoring (0.0-1.0 scale; >0.10 present, >0.30 strong topical focus) rather than keyword density. The operator question is no longer 'is my keyword density 1-2%' — it's 'does my page cover the entities and sub-topics a knowledgeable reader would expect on this topic?' If competitor top-10 pages all mention concepts your page doesn't, you have a topical-coverage gap, not a density gap.
How to do this
Audit:
Open the top 5 ranking pages for your target query.
List the named entities, sub-topics, and adjacent concepts each one covers.
Compare to your page. Note gaps.
For each gap: either you have a defensible reason to skip it (e.g. it's covered on a sister page; cluster cross-link) OR you should add a section.
Tools that do entity-coverage analysis: Surfer, Clearscope, Frase, MarketMuse, NeuronWriter. Use the ENTITY-COVERAGE column. Ignore the keyword-density column.
What NOT to do: stuff every competitor entity into your page just to chase coverage. Add coverage only where the entity is genuinely relevant to your page's specific frame.
Important Length matches the topic — write what the topic actually needs, never pad to a word count Word count is correlation, not causation. Surfer SEO: 'word count alone is not a ranking factor — Google's LLMs now evaluate meaning, relevance, and topical completeness rather than length.' Post-HCU absorption into core ranking makes padding worse than ever — Navboost tracks ~13 months of user-interaction signals (Gabe, December 2025) and users who bounce because the answer is buried under filler are an active negative.
How to do this
The 2026 operator stance:
Write the length the topic actually needs. Short for narrow answers, longer for genuinely complex topics.
Use entity-coverage tools to verify completeness, NOT to verify word count.
If your draft is shorter than competitor top-10 by 50%+: check whether you're missing genuine sub-topics, not whether you're hitting a word target.
If your draft is LONGER than competitor top-10 by 50%+: check honestly whether the extra is information density or padding.
Gabe's framing: 'Hell hath no fury like a user scorned.' Padding scorns the user; the user bounces; Navboost records the bounce; the page demotes. The shortest piece that completely answers the query wins more often than the longest one.
Important Periodic content cull — pages with low value get noindex or 410, not 'leave as-is' Leigh Stark's SEL HCU-recovery case study (October 2024): the highest-impact lever was content audit + culling underperformers — '10-20% were great, and the remainder was — not.' Underperforming pages were deleted or de-indexed. The HCU absorption into core (March 2024) means dead-weight pages drag down sitewide quality continuously — not just during an HCU re-run.
How to do this
Quarterly cull (or after any major core update):
Export GSC → Performance → Pages → sort by impressions (last 90 days).
Identify the bottom quartile by impressions AND clicks AND inbound links. These are candidates for cull.
For each: decide — (a) consolidate into a stronger page with a 301; (b) noindex if the page has a non-SEO purpose (e.g. login, internal); (c) return 410 Gone if the page is genuinely dead.
Never just leave low-value pages indexed but unattended — the sitewide quality signal compounds.
This is one of the few 'felt' on-page levers post-HCU. Stark's case study site's recovery was driven primarily by this + the brand/trust signals — NOT by backlinks.

AI Overviews / ChatGPT / Perplexity / Claude break pages into passages and retrieve passages, not pages. ~55% of AI Overview citations come from the top 30% of the page (SEL data). The page must be answer-first AND chunkable.

Critical Answer-first lead paragraph — 30-50 words answering the page's primary question in the first 1-2 sentences ~55% of AI Overview citations come from the top 30% of the page (Search Engine Land 'answer-first content' data). The traditional 'build up to the answer' essay style buries the citation target. Answer-first paragraphs are also the most-cited pattern in Perplexity + ChatGPT Search extractions. The lead 30-50 words are the single most valuable real estate on the page.
How to do this
Pattern:
State the page's primary answer in the first 1-2 sentences (30-50 words).
THEN expand with detail, nuance, qualifications.
The full intro can still be 200-400 words — but the LEAD must be answerable, not setup.
Operator test: if you cut everything except the first paragraph and showed it to a stranger, could they tell what the page answers? If no, rewrite.
Example bad lead: 'In today's fast-paced digital landscape, on-page SEO has become more important than ever as the AI era reshapes how content is discovered...'
Example good lead: 'The 2026 on-page SEO audit is 35 items across 7 categories. The single highest-leverage fix is making your title and H1 thematically coherent — that one change drops Google's title-rewrite rate on your page from ~76% to ~20%.'
Run the on-page audit free
Critical Logical H2/H3 hierarchy — each subheading frames a single concept independently understandable Aleyda Solís, July 2025: 'AI search engines don't index or retrieve whole pages; they break content into passages or chunks and retrieve the most relevant segments for synthesis.' Her optimization prescription: each section should be independently understandable without surrounding context, one concept per passage, clear HTML subheadings (H2/H3) for every subtopic. If a section needs context from the section before it, AI extraction misses the chunk.
How to do this
Audit each H2/H3 section:
Can the section stand alone? Copy a single H2 section to a blank document. Does it still answer a real question without referring backward?
Does the H2/H3 phrase the section as a single concept (ideally a question or a noun phrase)?
Is the H2 hierarchy logical — no H4 nested under an H2 with no H3 between them?
Common failure: an H2 like 'Step 3: configure the headers' that only makes sense if you read Steps 1-2 first. Fix: 'Configure response headers for SEO crawling' (self-explanatory).
The chunked-extraction test is the same as the AI-Overview-citation test. Pages that pass one tend to pass the other.
Important Paragraph rhythm: ~60-100 words per paragraph, sentences ≤ 20 words SEL's answer-first guide and operator consensus 2026: AI engines extract chunks of comprehensible prose; long monolithic paragraphs (200+ words) and run-on sentences (>30 words) reduce chunk quality. The 60-100-word paragraph and ≤20-word sentence rhythm is what's been shown to extract cleanly across ChatGPT, Claude, Perplexity, and Google AI Overview / AI Mode.
How to do this
Edit pass:
Any paragraph > 120 words: break into two.
Any sentence > 25 words: rewrite as two sentences.
Lists and tables count separately — they extract well, use them where data is comparable.
This is NOT a hard rule — the occasional longer paragraph or sentence is fine when the content demands it. The pattern matters, not the rigid count.
Operator test: read your page on mobile. If you find yourself scrolling through a wall of text with no visual rest, the paragraphs are too long for both human reading AND AI extraction.
Important Add citations, statistics, and direct quotations from authoritative sources where relevant Princeton GEO paper (Aggarwal et al., KDD 2024, peer-reviewed): adding citations, direct quotations from authoritative sources, and statistics increases source visibility in generative-engine responses by up to 40% across diverse queries and domains. This is the only peer-reviewed quantitative finding in the GEO literature — directional but well-evidenced. Pages with named-source density become 'secondary sources LLMs can quote-of-a-quote from'.
How to do this
For every long-form / pillar page, ensure:
At least 3-5 named-source citations per 1,000 words (e.g. 'Per the McAlpin / Search Engine Land study, Q1 2025...').
At least 1 directly-quoted statement from a named industry authority per 1,000 words.
At least 2 specific statistics with the source named (e.g. '76.04% of titles rewritten per McAlpin').
What NOT to do: cite-pad with low-authority sources just to hit a count. AI engines extract more reliably from pages that cite recognizable authoritative sources (Mueller, Illyes, named operators, peer-reviewed papers) than from pages that cite vendor blogs.
Use the brief / research-notes file as the source-of-truth for citations; never invent quotes.
Nice Tables and bullet lists for comparable data — LLMs extract structured data cleanly LLM extraction loves structured comparable data. A table comparing 3-5 options on 5 dimensions extracts as 15-25 discrete facts an AI engine can cite. The same data buried in prose extracts as 1-2 generalized statements. Bullet lists similarly outperform prose when the content is genuinely list-shaped (steps, options, features, criteria).
How to do this
Convert to tables when:
Comparing 2+ options on 2+ dimensions.
Presenting tier / pricing / spec data.
Showing before/after / right/wrong patterns.
Convert to bullet lists when:
Sequential steps (use ordered <ol>).
Independent options / criteria (use unordered <ul>).
Quick-reference cheat sheets.
Keep prose for arguments, narratives, nuanced explanations — that's where prose still wins. The 2026 win is matching structure to content shape, not 'use bullets for everything.'

Google has been actively retiring rich-result types (FAQ May 2026, HowTo 2023, 7 more June 2025). Ship the supported set, validate, audit quarterly.

Critical Article schema on every blog post / news article — required fields populated, JSON-LD only Article is one of the few schema types Google has NOT retired. Required fields: headline, image, datePublished, dateModified, author (with Person reference). Recommended: publisher (Organization), wordCount, mainEntityOfPage. JSON-LD is Google's official format recommendation; don't use Microdata or RDFa. The author reference is load-bearing for E-E-A-T (cross-link cq-01).
How to do this
Per-article checklist:
JSON-LD block in <head> or before closing </body>.
@type: Article (or BlogPosting / NewsArticle for specific subtypes).
headline: matches the page H1.
image: full URL, 1200×630 minimum for best surface coverage.
datePublished + dateModified: real dates, ISO 8601.
author: Person object with name, url pointing to author bio page.
publisher: Organization object with name + logo.
mainEntityOfPage: @id matching the page canonical URL.
Validate every block in Rich Results Test. The block should resolve cleanly with the Article eligibility flag.
Run the on-page audit free
Important Ship ONLY schema types Google's Rich Results docs list — skip Schema.org maximalism Google's own structured-data intro: 'rely on the Google Search Central documentation as definitive for Google Search behavior, rather than the schema.org documentation.' Schema.org defines hundreds of types Google doesn't surface as rich results. Schema maximalism is calorie burn + maintenance liability — the spec drifts, Google retired 7 more types in June 2025 alone (Book Actions, Course Info, Claim Review, Estimated Salary, Learning Video, Special Announcement, Vehicle Listing).
How to do this
The 2026 'ship-only-this' set:
Organization — homepage; with logo, url, sameAs cluster.
Article / BlogPosting / NewsArticle — blog/news pages.
BreadcrumbList — every page deeper than 1 level.
Person / ProfilePage — author bio pages.
Product — product pages (with offers + aggregateRating ONLY if real reviews exist).
LocalBusiness — physical-location businesses.
Recipe — recipe sites.
Video — video pages.
Event — event pages.
FAQPage — optional, only on real Q&A content (rich result gone May 2026; schema still valid).
Audit quarterly: paste your sitemap into GSC → Enhancements. Any newly-deprecated schema type will surface as warnings — strip those blocks.
Important Validate every schema block in Google's Rich Results Test BEFORE publishing Invalid JSON-LD (missing required fields, malformed @type, broken @id references) is silently dropped by Google. The schema looks correct in your CMS but earns zero rich-result eligibility. Validating in Rich Results Test catches this in 30 seconds per page — vs catching it in GSC's Enhancements report weeks later after rankings are already affected.
How to do this
Pre-publish workflow:
Open Rich Results Test.
Paste the staging URL OR raw HTML.
Confirm: 'Page is eligible for rich results' AND the specific feature you wanted (Article, Product, etc.) appears in the detected list.
Click each detected block; confirm no errors / warnings.
Common bugs:
Missing datePublished on Article (most common).
Image URL is a relative path instead of absolute.
Author Person reference uses a name string instead of an object.
Product missing brand or offers.
Multiple JSON-LD blocks for the same @type (template + plugin both emitting).
Run the on-page audit free
Critical Do NOT ship AggregateRating with fake review counts — manual-action risk Google has cracked down on inflated AggregateRating schema with manual actions through 2025-2026. LinkGuard's own removal of fake aggregateRating 4.9/127 from the landing page in April 2026 was prompted by exactly this risk. Google's structured-data guidelines require the rating to be visible to users on the page and reflect a genuine assessment of the item — fabricated counts fail both criteria and put your entire structured-data feature eligibility at risk, not just the rating.
How to do this
Two cases, two answers:
You have real, on-site, user-submitted reviews: ship AggregateRating with the actual count and average. Display it visibly on the page. Update as reviews accumulate.
You don't have real reviews yet: don't ship AggregateRating at all. Empty is fine. Lying is a manual-action risk that nukes ALL your rich-result eligibility, not just the ratings.
Same logic for Review schema (real reviewer names), Recipe ratings on commercial pages, Product aggregateRating without real reviews. The schema describes the page; the page describes the truth.

Alt text is now part of the AI-readability layer — vision-language models (CLIP, Gemini Vision, GPT-4o vision) detect objects, OCR on-image text, and fuse with alt + caption. Same fundamentals as 2018, sharper consequences in 2026.

Critical Every meaningful image has alt text — descriptive, accurate, ≤125 characters, no 'image of' prefix Alt text serves three audiences in 2026: screen readers (accessibility — 71.6% of screen-reader users navigate by headings per WebAIM), classical Google Image Search (still real organic surface for visual queries), and vision-language models (the new layer — CLIP / Gemini Vision / GPT-4o vision OCR on-image text and fuse with alt). Stuffed alt text actively hurts the VLM layer because it contradicts what the model sees in pixels.
How to do this
Alt text rules:
Descriptive of what's actually in the image (not the keyword you want to rank for).
≤125 characters, typically 5-15 words.
Never prefix with 'image of' / 'picture of' — screen readers already announce the image type.
Include the keyword ONLY if it's a natural part of the description.
Decorative images: alt="" (empty alt; screen readers skip).
Audit: Screaming Frog → Images → Missing Alt Text. Fix the meaningful images; explicitly mark the decorative ones with empty alt.
Example bad: alt="on page seo checklist best seo tips 2026 optimization guide" (stuffed; contradicts the actual image; works against VLM extraction).
Example good: alt="Screenshot of Google Search Console performance report showing 76% title rewrite rate Q1 2025" (describes the image, includes context, fits in 125 chars).
Run the on-page audit free
Nice Image filenames are descriptive kebab-case — not IMG_4521.jpg Image filenames are a weak ranking signal but a real one — Google uses them as one input to image-search relevance, and VLM-extraction tools weight filename + alt text + surrounding caption together. red-ceramic-mug-on-wood.jpg beats IMG_4521.jpg for both. Costs nothing if you fix at upload time.
How to do this
Rules:
Descriptive kebab-case: title-rewrite-rate-2025.png, not IMG_4521.JPG.
Lowercase, no spaces, no special characters except hyphens.
Short (3-6 words usually).
Matches the image content, not the keyword.
For existing images with bad filenames: don't rename retroactively unless you can also update every reference and 301 the old URLs (filename change = new URL). The cost is rarely worth it for indexed images. Fix at upload time going forward.
Important Modern image format (WebP / AVIF) with width + height + lazy loading Modern format (WebP / AVIF) is 25-50% smaller than JPEG / PNG at the same quality — direct LCP improvement. Explicit width and height attributes reserve layout space and prevent Cumulative Layout Shift (CLS). loading="lazy" defers below-fold images, freeing the LCP budget for the hero image. All three are 2026 Core Web Vitals table stakes.
How to do this
For every image:
<img src="/img/hero.webp" width="1200" height="675" alt="Descriptive alt text" loading="lazy">
Hero / LCP image (the largest above-the-fold image): exception — DO NOT lazy-load it. Add fetchpriority="high" and preload:
<link rel="preload" as="image" href="/img/hero.webp" fetchpriority="high">
Image optimization pipeline (one-time setup): use Sharp / Squoosh / ImageOptim / your CDN's auto-format to convert at upload, serve WebP with JPEG fallback via <picture>. For CMS-uploaded images, configure the CMS to do this automatically.
Nice Captions used where they add real context — captions are read more than body copy near images David Ogilvy's classic finding (still cited in 2026 editorial-design literature): captions get ~3x the readership of body copy near the image. For editorial / long-form pages with original photography or data visualizations, a caption that names what's in the image AND why it matters is one of the highest-leverage 100-word edits available. Captions are also extracted by VLM tools alongside alt text.
How to do this
Add a caption when:
The image is original (photo, screenshot, chart) and not self-explanatory.
The image presents data the reader will benefit from understanding the source / methodology of.
The image illustrates a concept that benefits from a one-line explanation.
Skip the caption when:
The image is purely decorative.
The image is self-explanatory (e.g. an obvious illustration matching the section's text).
Caption format: short (≤25 words), italic or visually distinct from body copy, names the source if the image is data. Don't repeat the alt text verbatim — alt text is for screen readers, the caption is for sighted readers.

One of the few on-page levers with measured-correlation impact on organic traffic (Cyrus Shepard / Zyppy 23M-link study). The mechanism is anchor-text variety, not raw link count.

Important Every indexable page receives 3-40 contextual internal links — flatlines beyond ~40-45 Cyrus Shepard / Zyppy analyzed ~23 million internal links across ~1,800 websites, ~520,000 URLs (updated Feb 2026). Traffic rose as internal links to a page increased — but only up to ~40-45 links per page. Beyond that, traffic flatlined or declined. Below 3 internal links, the page is structurally orphaned. The 3-40 contextual-link window is the operator-safe zone.
How to do this
Audit:
Crawl your site with Screaming Frog → 'Inlinks' for each priority URL.
Flag pages with < 3 inbound internal links (orphans / near-orphans).
Flag pages with > 50 inbound internal links (probably over-linked from sitewide elements like footer / sidebar).
For orphans: add contextual links from related pillar/cluster pages.
For over-linked: consider removing the sitewide footer/sidebar link if the page doesn't deserve that prominence.
'Contextual' means links from within body content, not from sitewide navigation / footer. Sitewide nav links count for crawlability; contextual body links count for ranking signal.
Critical Anchor text on internal links is VARIED — same exact-match phrase repeated dilutes the signal Cyrus Shepard's Zyppy 23M-link study: anchor-text variety was the driving correlated factor with traffic, NOT raw link count. Up to 5x traffic uplift from better anchor variety. Recent Zyppy update flags that pathologically high anchor diversity may also struggle under recent algorithm updates — so the operator stance is 'varied + relevant + natural,' not 'manufactured variety for variety's sake.'
How to do this
For each priority page, audit its INTERNAL anchor text portfolio:
Crawl your site, filter for internal links pointing at the target URL.
List all anchor texts pointing to that URL.
If the same exact-match phrase repeats > 40% of the time: dilute it — rewrite 2-3 links to use partial-match / descriptive variants.
If anchors are highly varied but unnatural / forced: consolidate toward 2-3 natural phrasings.
Healthy distribution: 30-50% branded / descriptive ('our pricing checklist'), 20-30% partial-match ('the 2026 pricing checklist'), <10% exact-match ('pricing checklist'). Internal portfolios usually skew MORE branded than external in practice (content people default to 'our X' / 'the X checklist'); see our anchor-text portfolio checklist for the external-side framework, which uses a similar but typically more aggressive distribution.
Audit anchor distribution free
Critical Internal links point to FINAL URLs — never into a 301 redirect chain In a typical client audit this is in the top 3 most-common findings AND the cheapest to fix — template-level search-and-replace, one afternoon. Every internal link that points at a URL which 301-redirects is one extra hop the user (and Googlebot) has to follow. Crawl budget tax. Latency tax. The 301s handle external referrers you can't update; internal links you control should point directly at the final destination.
How to do this
Audit (Screaming Frog → Internal Links → filter Status Code 301):
For each internal link that 301-redirects, update the source HTML to point at the final destination.
Re-crawl to confirm no internal links land on 301 / 302.
This is the cheapest internal-SEO win available — usually a 1-hour search-and-replace across templates. Body-content links can usually be fixed in the CMS editor; template-level links (footer, menu, sidebar) need a dev hand-off.
Bonus: same audit catches any internal links to 404s (filter Status Code 404). Fix those by either restoring the target URL OR removing the link.
Important Pillar / cluster structure — pillar covers topic comprehensively; cluster pages cover facets and link back Aleyda Solís's 2025 framing: topical authority requires a pillar page that covers a topic comprehensively, with cluster pages covering specific facets, all internal-linked back to the pillar. This is the on-page structure that supports passage retrieval — AI engines extract chunks from cluster pages AND attribute the topical authority to the pillar. Without pillar/cluster organization, the same content scattered across 30 unconnected pages performs ~30-50% worse.
How to do this
For each of your top 3-5 topics, audit:
Is there ONE clearly identifiable pillar page that covers the topic comprehensively?
Does the pillar internal-link to all the cluster pages?
Does each cluster page internal-link BACK to the pillar?
Do cluster pages internal-link to RELATED cluster pages (not just the pillar)?
If your top topic has no pillar: write one. If your pillar exists but doesn't link to cluster pages: add the links. If cluster pages don't link back: add the links (usually a 'See also' / 'Related reading' block at the end).
This is the structural change that compounds over months — internal-link signal aggregation lifts the entire cluster's ranking, not just the pillar.

Five things vendor blogs and SEO course-sellers currently sell as on-page best practice that don't survive scrutiny against 2025-2026 evidence. If you've been paying for any of these, you're not behind — you're early enough to redirect.

Important Don't optimize for keyword density — Mueller has debunked it since 2011; modern Google uses NLP entity-salience John Mueller has confirmed keyword density is not a ranking factor as early as 2011 and has restated it multiple times since. Modern Google's relevance assessment uses NLP entity-salience scoring (0.0-1.0 scale; >0.10 present, >0.30 strong topical focus), not term-frequency-density. Vendor tools that still ship a density column do so for inertia, not evidence. Operator metric for 2026: entity coverage / topical completeness, not density percentage.

How to do this

What's true: keyword density as a number you optimize toward (e.g. 'aim for 1-2%') stopped mattering for ranking ~14 years ago.
What replaced it: entity-coverage / topical-completeness — does your page mention the entities a knowledgeable reader would expect on this topic, with appropriate prominence (primary entity above 0.30 salience, supporting entities present at all)?
Tools that report keyword density: ignore that column. Tools that report entity-coverage gaps (Surfer, Clearscope, Frase, MarketMuse, NeuronWriter): use that column.
If your team's content brief has a 'target keyword density' field: delete it. Replace with 'entities the page must cover' as a checklist.
Critical Don't stuff exact-match keyword anchors on internal links — same anti-over-optimization signals as external Exact-match anchor over-optimization triggers the same algorithmic detection (legacy Penguin signals folded into core ranking) on INTERNAL links as on external. Cyrus Shepard / Zyppy 23M-link study: anchor-text variety correlated with traffic uplift, NOT raw repetition. Operator-safe ceiling: exact-match keyword anchors ≤ 5-10% of any one page's internal anchor profile.
How to do this
Audit per-page anchor distribution:
Crawl your site, filter internal links to the target URL, list all anchor texts.
Calculate the share of EXACT-MATCH keyword anchors (e.g. 'on-page SEO checklist' if that's your target keyword).
If > 10%: dilute by rewriting some anchors to descriptive / partial-match / branded variants.
Healthy 2026 internal-anchor balance:
30-50% branded / descriptive ('our on-page checklist', 'the 2026 audit guide').
20-30% partial-match ('on-page SEO audit checklist', 'on-page checklist for 2026').
<10% exact-match ('on-page SEO checklist').
Audit anchor distribution free
Important Don't ship every schema type Schema.org defines — Google has been actively RETIRING rich-result features Google's own structured-data intro: 'rely on the Google Search Central documentation as definitive for Google Search behavior, rather than the schema.org documentation.' In 2025-2026 Google has been actively retiring rich-result types — FAQ rich result killed May 2026, HowTo killed 2023-2024, seven more types retired in June 2025 (Book Actions, Course Info, Claim Review, Estimated Salary, Learning Video, Special Announcement, Vehicle Listing). Maximalist schema is now a maintenance liability — the spec is shrinking, not growing.
How to do this
Don't ship:
Schema types Google's Rich Results docs don't currently support for your page type.
FAQ schema purely to chase the (gone) SERP feature.
HowTo schema for the (gone) SERP feature.
Any of the 7 types retired June 2025 (unless they survive for ad attribution / structured-data validators in other systems).
What to do quarterly:
Review Google's structured-data search gallery — anything you ship that's no longer listed: strip.
Check GSC → Enhancements → any new 'invalid' / 'no longer supported' warnings: address.
Per Google: unused/deprecated schema doesn't actively harm rankings. But maintaining schema you don't need is calorie burn — and a deprecation that turns into a validator warning can erode trust in the rest of your structured data.
Important Don't pad content to a word-count target — Navboost punishes the bounce from padding Word count is correlation, not causation. Surfer SEO: 'word count alone is not a ranking factor.' Post-HCU absorption into core ranking (March 2024) makes this worse than ever — Navboost tracks ~13 months of user-interaction signals (Gabe, December 2025) and users who bounce because the answer is buried under filler are an active negative regardless of whether the page is 800 or 2,500 words.
How to do this
The 2026 operator stance:
Write the length the topic actually needs.
Use entity-coverage tools to verify completeness, NOT to verify length.
If your draft hits the target topic completely in 800 words, ship 800 words.
If the topic genuinely requires 2,500 words for completeness, ship 2,500 — but the 2,500 must be 2,500 of substance, not 1,200 of substance + 1,300 of padding.
Operator test (run on your own page): can you delete any paragraph without losing a fact or argument the reader needs? If yes, delete it. Padding scorns the user; the user bounces; Navboost records the bounce; the page demotes.
Critical Don't publish mass AI-generated content with no human gate — scaled-content-abuse manual-action risk since June 2025 Google's policy: AI content as a category is NOT penalized; scaled content abuse IS. Manual actions for scaled-content-abuse rolled out around June 2025, targeting mass-produced AI content without unique value. Glenn Gabe, December 2025: 'publishing AI content without human intervention is extremely risky.' Pattern-based detection (volume + thinness + low originality + scaled production), not an AI-detection threshold.

How to do this

Safe pattern: AI assists, human gates. Drafts, outlines, transcript cleanup — fine. AI fact-checks, original sources, named author bylines all stay human.
Unsafe pattern: LLM-to-publish with no human editor. Scaled production (e.g. 100+ pages per week from one workflow) with low originality.
Operator test: if your average human edit time per 1,500-word AI-assisted article is < 20 minutes, you're in the policy zone whether the editor signed off or not. Either raise edit time or reduce volume.
Recovery from a scaled-content-abuse manual action takes months, requires deleting the offending pages, and (per Gabe's recovery framing) only works if Google sees the cleanup sustained over time. The cost-saving isn't worth the risk.

About the Author

Andrei

SEO and digital marketing professional with 13+ years of experience. Started as a website administrator in 2011, transitioned to SEO, and achieved top-3 rankings for competitive keywords. Co-founded a consulting firm specializing in marketing audits for companies in Ukraine and internationally. Built LinkGuard to solve the problem he experienced firsthand: most SEO teams purchase links but never monitor their survival. Based in Kyiv, Ukraine.