GEO Red Flags: The Claims That Don't Hold Up, and the Sources That Say So

99Visibility Research · Updated October 21, 2026 · 8 min read

Short answer

Most GEO red flags share one trait: they promise an outcome the evidence does not support. A 40% visibility uplift, an 'AI rank' in Google, schema that boosts citations, llms.txt as a requirement, and an E-E-A-T score to improve are each contradicted by a primary source. The claims that hold up are about access: whether AI crawlers can reach, read and index your pages.

Key takeaways

Nearly every published GEO effect size was measured on content already inside the model's context, so it says little about whether AI engines find you in the first place.
Google states structured data isn't required for its generative AI features, and Ahrefs' matched test found AI Overview citations fell 4.6% after pages added schema.
There is no E-E-A-T score, no Google AI rank, and no engine that computes domain authority, so any tool reporting one invented it.
The honest caveats matter: existing schema does no harm, and llms.txt is ignored by Google Search, not proven useless everywhere.
The levers with real evidence are subtractive: server-side rendering, per-bot robots.txt rules, firewall reachability, index eligibility and snippet controls.

Why most GEO claims fail the same test

An AI answer passes through four stages. Parametric knowledge is what the model learned in training. Access is whether a crawler can fetch and store your page. Retrieval is whether your page enters the model's context for a given question. Generation is which of those retrieved pages the model uses and cites.

Almost every GEO technique with a published effect size was tested at the last stage, on documents researchers had already placed in the model's context. A critical survey of 45 GEO studies (arXiv, July 2026) concluded that already-retrieved content can change its citation, but no reviewed technique shows a stable, cross-platform effect on discoverability. C-SEO Bench (NeurIPS 2025) went further: most conversational SEO methods were largely ineffective and often hurt ranking inside the context.

What that means for you: a vendor quoting a lab number is usually quoting a generation-stage result as if it were a discoverability outcome. Your first question for any claim should be which stage it was measured at. If you want the broader argument, see whether GEO is legit or just hype.

The red flags and the sources that contradict them

Use this table as a checklist when you read a proposal, a tool report or a LinkedIn post.

Claim you'll hear	What the primary source says	Honest caveat
GEO gets you up to 40% more visibility	The GEO paper tested an engine the authors built, with sources supplied in context	It is a real lab finding about content already retrieved
Track your AI rank in Google	Search Console's generative AI report shows impressions only, with no position, clicks or CTR	Presence and citation counts can be measured, if retrieval state is disclosed
Add schema to get cited	Google says structured data isn't required for generative AI search; Ahrefs found AI Overview citations fell 4.6%	Ahrefs only tested pages already heavily cited; entry into the answer is untested
Add FAQ schema for AI answers	Google deprecated FAQ rich results in May 2026	Q&A content on the page is a separate thing from the markup
Publish an llms.txt	Google says Search ignores it; no other engine documents reading it	Google says it's fine to keep one for other systems
Improve your E-E-A-T score	Google says E-E-A-T 'isn't a specific ranking factor'	Named authors and cited sources can still be good practice
Build domain authority	Domain Authority is a Moz metric no engine computes	Google does publish a news-scoped topic authority system
Allow GPTBot so ChatGPT finds you	GPTBot governs training only; OAI-SearchBot governs ChatGPT search	Blocking GPTBot doesn't remove you from ChatGPT search either
Chunk content or write AI-friendly prose	Google rejects both and says there is no ideal page length	Chunking research measures segmentation done by the engine, not you
Ping IndexNow and AI cites you faster	The spec never mentions AI; a 200 response confirms receipt only	Bing consumes IndexNow; any AI effect is unmeasured

The 40% uplift and the AI rank, up close

The 40% figure

The number traces to the original GEO paper (Aggarwal et al., KDD 2024). The authors built a two-step engine and supplied the sources in context. Even their one Perplexity test uploaded source text as files, because Perplexity did not let users specify source URLs. The main metric counted words the intervention itself added, and the paper reports a 30 to 40% band across three methods, not the itemised figures that circulate. Nothing in it measures whether engines discover your pages. The full breakdown is in the 40% GEO visibility claim, examined.

The AI rank

Google's first-party generative AI report exposes impressions. It has no position dimension, so there is no ground truth to check any 'AI rank' against. Google also says AI Overviews and AI Mode 'may use different models and techniques' and can show different links. On other platforms, retrieval fires at the model's discretion per question, so any single score averages answers given with and without live search. That includes every vendor's score, ours too.

The next move: when a report shows a rank, ask what it was validated against. When an agency promises a position or a citation, read whether to trust an agency that guarantees AI citations before you sign.

Schema, llms.txt and the scores nobody computes

Schema and FAQ schema

Google states structured data isn't required for generative AI search, and no AI-specific schema type exists. Ahrefs tracked 1,885 pages that added schema against matched controls and found no major uplift on any platform, with AI Overview citations falling 4.6%. FAQ rich results were deprecated in May 2026.

The caveat cuts both ways. Every page Ahrefs treated already had more than 100 AI Overview citations, so whether schema helps a page with zero citations enter the answer is untested, not refuted. Google also says existing markup does no harm. The honest line: schema is not a lever for AI answers.

llms.txt

Google's AI optimization guide says Search ignores AI text files, and adding one 'will neither harm nor help' your visibility in Google Search. No other engine documents parsing it. The statement is Google-scoped, so the accurate claim is 'no evidence it works', not 'proven useless everywhere'.

E-E-A-T and domain authority

Google says E-E-A-T 'isn't a specific ranking factor'. There is no number to move. Domain Authority is a third-party regression built from backlinks and traffic, and Google publishes no site-level authority score. If a tool shows you either score, it made it up.

If these tasks are already on your roadmap, retiring unsupported GEO tasks shows how to wind them down without breaking anything.

What is in your control and what is not

In your control

Rendering. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. Content that appears only after JavaScript runs may be invisible to them. Googlebot does render.
Per-bot robots.txt rules. OpenAI says 'Each setting is independent of the others'. OAI-SearchBot, PerplexityBot and Claude-SearchBot are the search crawlers to check.
Network reachability. A firewall or CDN can block an AI crawler that robots.txt allows. OpenAI says eligibility needs both the robots.txt permission and a host that accepts its published crawler IPs.
Index and snippet eligibility. Google says a page must be indexed and eligible to show with a snippet to appear in AI Overviews or AI Mode, with 'no additional technical requirements'.
Snippet controls. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out of AI answers, which matters when a template applies them by accident.
Plain facts. Stating a price, a number or a one-line definition on a fetchable page gives an engine a sentence it can quote.

Not in your control

Whether a given question triggers live retrieval at all.
Your position in the model's candidate list.
What the model learned before its training cutoff.
How the model words or cites its answer.

Being named on the sites engines already cite shows a strong association with appearing in answers. It is an association, not a promised effect.

A practical order of work

Work through access before anything else, because every later step depends on it.

Check robots.txt per bot. Confirm the search crawlers are allowed. Decide on training crawlers separately. A rule like this keeps you in ChatGPT search while opting out of training:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

Request your pages as each AI crawler. A 403 or a challenge page means your firewall is turning the crawler away, whatever robots.txt says. The AI Crawler Checker runs this request for each crawler and shows which ones get blocked.
Diff raw HTML against the rendered page. Anything important that appears only after JavaScript runs needs server-side rendering.
Audit snippet controls. Search templates for stray nosnippet or data-nosnippet attributes.
Confirm index status engine by engine. Check Google Search Console and Bing Webmaster Tools separately.
Put key facts in plain text. Make pricing, product names and definitions consistent across every page.
Then retire or pause the unsupported tasks from the table above.

To turn steps 1 to 4 into work your engineers will pick up, use developer tickets for AI access fixes.

Common mistakes when you spot a red flag

Overcorrecting. 'Schema is useless, remove it' is not supported either. Google says existing markup does no harm. Leave it and stop counting it as AI work.
Throwing out a vendor's whole plan. A proposal with one bad claim may still include sound access fixes. Keep those.
Trusting a single blended score. One engine can cite you heavily while others barely do. Ask for results per platform, with retrieval state disclosed.
Treating case studies as proof. A before-and-after on one site, with no control, cannot separate the tactic from brand size or timing. See how to read GEO case studies with big citation gains.
Reading crawler visits as visibility. A training crawler visiting your site sends no referrals by design. If you're puzzled by the data, read why AI crawler visits don't show up in Google Analytics.
Hedging into mush. The evidence on access is strong. When your firewall blocks OAI-SearchBot, say so plainly and fix it.

Frequently asked questions

Are AI SEO services scams?

Not by default. Many services bundle real access work, like rendering and crawler permissions, with unsupported claims like schema boosts or E-E-A-T scores. The red flag is a promised outcome: a guaranteed citation, a rank, or a fixed percentage uplift. Nobody controls what the models say. A credible service tells you which of its numbers are measurements and which are its own model, and it fixes what blocks you rather than promising what the engines will do.

Should I remove schema markup I already have?

No. Google says existing structured data does no harm, and it still serves classic search features. The evidence only says schema is not a lever for AI answers: Ahrefs found AI Overview citations fell 4.6% on pages that added it, among pages already heavily cited. Keep what you have, stop reporting it as AI visibility work, and spend new effort on access fixes instead.

Does publishing an llms.txt file hurt my site?

Google says adding one will neither harm nor help your visibility in Google Search, because Search ignores it. Google also says it's fine to maintain one for other systems that use these files. No other major AI engine documents reading it. So it is harmless but unproven. Don't let it displace work that has evidence behind it, like server-side rendering of your key facts.

What should I ask a GEO vendor before hiring them?

Ask which stage each claimed result was measured at: access, retrieval or generation. Ask what any rank or score was validated against. Ask whether results are reported per platform and whether live search was on. Ask for the primary source behind each recommendation. If the answers point to lab studies on content already in context, or to scores no engine computes, treat the pitch as a red flag.

Sources

This page is AI-assisted. It was written from a verified evidence base and checked against our editorial policy. Engine behaviour changes often; facts are as of October 21, 2026.