Updated
Updated What changed on
Fact-checked The AI crawler names and robots.txt behaviour, Google-Extended, Google's guidance on chunking and structured data, and the Ahrefs ranking study were re-checked against the 14 October 2026 source checks. The schema citation studies (Growth Marshal February 2026, Ahrefs May 2026, SALT.agency September 2025) and Google's AI optimisation guide were also opened and re-checked on 14 October 2026. Nothing else on the page was re-checked.
- Split the AI crawlers into search crawlers (OAI-SearchBot, Claude-SearchBot, PerplexityBot), training crawlers (GPTBot, ClaudeBot) and user-triggered fetchers (ChatGPT-User, Claude-User, Perplexity-User). GPTBot and ClaudeBot only control training.
- Replaced anthropic-ai with Claude-User, the bot Anthropic actually lists, and updated the OpenAI, Anthropic and Perplexity crawler links to their current pages.
- Corrected the robots.txt advice: OpenAI, Google, Perplexity and Meta say their user-triggered fetchers may ignore robots.txt.
- Corrected Google-Extended: it is a robots.txt token for Gemini training and grounding, does not control AI Overviews and never shows in your logs.
- Removed the 100-200 word passage rule: Google's May 2026 guide says there is no need to chunk content for Google Search and no ideal page length.
- Corrected the schema advice: Google says structured data is not required for its AI features, and FAQ and How-to rich results no longer show in Google Search.
- Replaced an unsupported "30x" ranking claim with what the Ahrefs July 2025 study measured: a Spearman correlation of 0.347 between top-10 ranking and top-three AI Overview citation.
- Put the 61.7% vs 41.6% schema figures in context: they come from one February 2026 study of 730 ChatGPT and Gemini citations, compare Product and Review schema with generic types, and sit next to Ahrefs' May 2026 test, which found no citation lift from adding JSON-LD.
- Replaced the schema stat card with the Ahrefs May 2026 result and removed a "1,006 pages" sample size the study does not state.
- Corrected the sameAs figure: the "fewer than 4%" finding comes from the February 2026 study, not SALT.agency, whose 107,352-URL AI Mode analysis found no schema type beyond the standard set gave an advantage.
What does a GEO audit actually include?
The 30-second answer
Eight areas: AI bot access audit, schema audit, content extractability review, named-author/Person schema check, baseline citation reading against priority queries, off-site authority review, ranking foundations on AIO-triggering queries, engine-specific priority recommendations based on your buyer mix. Deliverable is a written audit document plus baseline citation report plus 60-day implementation plan. Mid-market scope takes 2-3 days of consultant time. Anyone delivering a "GEO audit" in under a day is doing a schema check, not an audit.
What an audit is measuring against
of URLs cited in Google AI Overviews appeared in the first 10 results blocks for that query. Ahrefs counts blocks, not organic results, so this is not the same as ranking in the top 10.
How it was measured: 863,000 keyword SERPs and 4 million AI Overview URLs. Ahrefs measures the top 10 as blocks, counting ads, featured snippets and video packs as separate blocks.
was Ahrefs' July 2025 figure. It counted only the top 3 citations per AI Overview, and Ahrefs has since changed its parsing, so the two years are not like for like.
How it was measured: 1.9 million citations from 1 million AI Overviews, analysing only the top 3 most visible citations in each one. Ahrefs notes it improved its parsing between this study and the 2026 update. Both differences mean the fall to 37.9% cannot be read as a like for like change.
of Google searches trigger an AI Overview, depending entirely on whose panel you read.
How it was measured: Pew measured 18% on real-user browsing data. Ahrefs measured 48% of its keyword set in March 2026. Similarweb reported 43%, up from 15% a year earlier, via TechCrunch on 27 July 2026. The panels sample different query mixes, so the spread is a methodology artefact. Anyone quoting a single figure as the number is overstating what is known.
Sources disagree on this one. The range is the honest answer.
Four numbers that frame the audit
Before the eight areas, the data that explains why the audit exists in the shape it does. None of this is from my own work; all four are published research from people who run citation panels at scale.
That last one is why the audit ranks prompt-position, not just presence. Showing up tenth in a Perplexity answer is close to invisible.
The eight areas
1AI bot access audit
GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, Claude-User, PerplexityBot, Perplexity-User, Google-Extended, AppleBot-Extended, MistralAI-User. Training crawlers (GPTBot, ClaudeBot) feed model knowledge. Search crawlers (OAI-SearchBot, Claude-SearchBot, PerplexityBot) build the engines' search indexes, and user-triggered fetchers (ChatGPT-User, Claude-User, Perplexity-User) fetch live during user prompts. Google-Extended is a robots.txt token for Gemini training and grounding; it does not control AI Overviews and never shows in your logs. Training and search crawlers follow robots.txt, but OpenAI, Google, Perplexity and Meta say their user-triggered fetchers may ignore it. You can still block training and allow search, which is the right default for most B2B sites. Test robots.txt rules and Cloudflare/CDN bot-fight rules. The most common silent block sits at the CDN, not the file.
2Schema audit
Article, Person, Organization, ProfessionalService or Service, BreadcrumbList. Google says structured data is not required for its AI features, but it still helps pages qualify for rich results. FAQ and How-to rich results no longer show in Google Search. The research on schema and AI citation is mixed, so the audit treats markup as hygiene, not a citation lever. A February 2026 study first published by Growth Marshal looked at 730 citations from ChatGPT and Gemini. Pages with Product or Review schema carrying prices, ratings and specifications were cited 61.7% of the time, against 41.6% for generic types such as Article, Organization and BreadcrumbList, and 59.8% for pages with no schema. Its own model found no significant effect from schema presence alone. Ahrefs' May 2026 test tracked 1,885 pages, drawn from URLs already cited in AI Overviews, that added JSON-LD, against 4,000 matched controls. It found no citation lift in AI Mode or ChatGPT, and a small AI Overview decline it could not pin on schema. SALT.agency's analysis of 107,352 URLs cited in AI Mode found no schema type beyond the standard set gave a measurable advantage. The same February 2026 study found fewer than 4% of schema-present pages used entity links such as Wikidata sameAs, too few for that study to test whether they help.
3Content extractability review
Plain-answer leads above the fold of each priority section. Google's May 2026 guide says that, for Google Search, there is no need to chunk content. Short paragraphs. Tables and lists for comparison and list content. Server-side or static rendering, because GPTBot, ClaudeBot and PerplexityBot do not execute JavaScript - 50-80% of content on client-rendered SPAs is invisible to them. The structural changes that decide whether the generation stage of RAG cites you or a cleaner source.
4Named-author + Person schema check
Author bylines on every content piece. Person schema linking to credentials and professional profiles. sameAs graph populated (Wikidata, LinkedIn, GitHub where applicable). Google says structured data is not required for its AI features, but it still helps pages qualify for rich results.
5Baseline citation reading
10-20 priority buyer queries. 3-5 named competitors. 5 engines (ChatGPT, AI Overviews, Perplexity, Claude, Gemini). 3 samples per query to smooth output variance. Citation frequency + share-of-voice + prompt-position + per-engine breakdown established. Without this baseline the audit has no measurement plane and you cannot tell three months later whether anything moved.
ChatGPT
AI Overviews
Perplexity
Claude
Gemini
6Off-site authority review
Wikipedia and Wikidata entity check. Knowledge Graph entity strength. Citations from credible publications. Reddit presence on the subreddits the engines actually surface. The signals that compound slowly but feed citation-pool eligibility.
7Ranking foundations on AIO-triggering queries
Ahrefs (July 2025) measured a Spearman correlation of 0.347 between ranking in the top 10 and being one of the top three AI Overview citations, and pages ranking first appeared in the top three citations only about half the time. The audit identifies which AIO-triggering queries your priority pages are NOT in the top 10 on, because that is where the rank improvement work pays back fastest for AI citation.
8Engine-specific priority recommendations
The audit closes with a recommended engine priority order for your buyer mix. UK mainstream B2B usually: ChatGPT + AI Overviews first. Technical and research-heavy buyers: add Perplexity + Claude. EU enterprise: add Mistral Le Chat. Without this, the audit hands you a generic checklist that wastes 30% of the budget on engines your buyers do not use.
How a GEO audit differs from an SEO audit
Same plumbing in places, completely different question being asked.
| Dimension | SEO audit | GEO audit |
|---|---|---|
| Question | Can Google find, index and rank this page? | When an AI engine answers this buyer question, does this page get cited? |
| Crawlers tested | Googlebot, Bingbot | Add OAI-SearchBot, Claude-SearchBot and PerplexityBot (search) and GPTBot and ClaudeBot (training); check Google-Extended as a robots.txt token, since it never shows in logs |
| Success metric | Rankings, organic clicks, impressions | Citation frequency, share-of-voice, prompt-position |
| Content shape | Long-form depth, internal linking | For Google Search, no chunking needed and no ideal page length (Google, May 2026); clear answers for other engines |
| Rendering | Google renders JS via WRS | Most AI crawlers do not render JS; raw HTML or nothing |
| Authority signal | Backlinks, domain rating | Wikipedia/Wikidata entity strength, named author, sameAs graph |
| Schema role | Hygiene for Rich Results | Checked for entity clarity, not as a citation lever: Google says it is not required for its AI features, and Ahrefs (May 2026) found no citation lift from adding JSON-LD |
| Output | Position changes over time | Per-engine citation report against named competitors |
They overlap on basics (fast pages, clean HTML, sensible structure). They diverge on everything that decides whether you get quoted instead of just ranked.
The metrics the audit actually tracks
Five numbers, measured at baseline and re-measured every quarter.
Bonus tracking when relevant: top-10 ranking coverage on AIO-triggering queries (Ahrefs, July 2025: Spearman correlation 0.347 between top-10 ranking and top-three AI Overview citation), and brand-mention sentiment in cited contexts.
The workflow, end to end
Scope the buyer queries
10-20 questions your actual buyers ask AI engines. Not keyword lists, not search-volume reports. Sales-call transcripts, support tickets, demo notes.
Run the baseline citation reading
Five engines, three samples per query, named competitor set. Output: a spreadsheet you can rank against in 90 days.
Bot access + technical pass
robots.txt, Cloudflare bot rules, JavaScript rendering check on priority pages, schema validation, author + Person + Organization graph.
Content extractability pass
Walk each priority page. Mark the 30-second answer block, tables and lists. Flag pages where the answer is buried in paragraph four.
Off-site + ranking review
Wikipedia/Wikidata entity check, Knowledge Graph strength, AIO-triggering queries where the page is not yet top-10.
Engine-priority recommendation
Score the buyer mix. Pick the two-or-three engines that matter. Write the implementation order so 70% of effort lands where 70% of citations will come from.
Write the document + 60-day plan
One section per audit area, current state + gap + actions. Sequenced by impact-vs-effort, not by audit area order.
What the deliverable looks like
The audit is the diagnostic; the fixes follow the structure in my GEO field guide, so the document is built to plug straight into that programme.
- Written audit document, one section per area above, current state + gap + prioritised actions
- Baseline citation spreadsheet (citation frequency + share-of-voice + prompt-position + per-engine breakdown)
- 60-day implementation plan sequenced by impact-vs-effort
- Engine-specific priority order for your buyer mix
- Re-measurement protocol so the quarter-three reading is comparable to the baseline
What is NOT a GEO audit
- A schema check delivered in 4 hours and called an audit
- A one-engine citation snapshot with no competitor comparison or baseline
- An llms.txt "audit" that just checks whether one exists (per SE Ranking's 300,000-domain study, llms.txt presence shows no measurable correlation with citation rate yet)
- A "GEO audit" that does not include a baseline citation reading against named competitors - that is the load-bearing input everything else depends on
- An AI-tool-generated PDF with no engine-specific recommendation at the end
More on GEO audits.
What is in a proper GEO audit?
Eight areas: AI bot access audit, schema audit, content extractability review, named-author/Person schema check, baseline citation reading against priority queries, off-site authority review, ranking foundations on AIO-triggering queries, engine-specific priority recommendations based on your buyer mix.
How is a GEO audit different from an SEO audit?
An SEO audit asks "can Google find, index and rank this page". A GEO audit asks "when an AI engine answers the buyer question, does this page get cited". Different success metric (citations and share-of-voice, not blue-link rankings), different crawlers (OAI-SearchBot, Claude-SearchBot and PerplexityBot for AI search, GPTBot and ClaudeBot for training, alongside Googlebot), different content shape (Google's May 2026 guide says that, for Google Search, there is no need to chunk content and there is no ideal page length), different scoring (per-engine citation frequency across five engines, not one results page).
How long should a GEO audit take?
For a mid-market site (50-200 pages), 2-3 days of consultant time. Larger sites or multi-region scopes take longer. Anyone delivering a "GEO audit" in under a day is checking schema and stopping; ask what the other six areas covered.
What is the deliverable?
A written audit document covering each of the eight areas with current state, gap analysis, and prioritised action list. Plus a baseline citation report against named competitors. Plus a 60-day implementation plan sequenced by impact-vs-effort.
Which metrics does a GEO audit actually measure?
Citation frequency (how often you are quoted per query), share-of-voice against named competitors, prompt-position (first recommendation versus tenth - AirOps 2026 data puts the first pick at 74% of user choice), per-engine breakdown (ChatGPT vs AI Overviews vs Perplexity vs Claude vs Gemini), entity recognition accuracy, and top-10 ranking coverage on AIO-triggering queries.
Do I need a separate GEO audit if I already have a recent SEO audit?
Yes. An SEO audit will not check AI search crawler access separately from training crawlers, and will not measure citation frequency in ChatGPT or Perplexity. Google says structured data is not required for its AI features, but it still helps pages qualify for rich results. Roughly 30% of the work overlaps; 70% is new ground.
Jason Burns
Independent UK SEO, GEO and AI consultant. In search since 2007. Portfolio includes 3M, BlackRock, Unilever and E.ON. Owner of SEO Moves Ltd since 2014.
More about Jason →Want a proper GEO audit on your site?
Send your site URL and your three biggest competitors. I scope out the audit in a one-page reply with a fixed price - no price-on-application game.
Request the audit