100 GEO audits. one drop of insights: why AI knows your brand but never recommends it
We're 009. Over the past months we ran 100 GEO audits across Web3, crypto, iGaming and B2B services. To mark audit #100, we're dropping the insights: the five mistakes that repeat from audit to audit, the myths the market is burning money on, and the real state of LLM visibility in every niche.
Dima Lizanets, strategic advisory at 009 — led the 100 GEO audits behind this report
published july 17, 2026 · 12 min read
chatgpt · claude · gemini · perplexity · grok · deepseek
- same mistakes for everyone?No. Roughly half of the issues in any single audit are unique to that project — its stack, its history, its niche. This report covers the other half: the mistakes that repeat everywhere.
- will AI recommend me?Almost certainly not: across 100 audits, only 4.5% of "who should I pick" answers named the audited brand — even though AI knows 87% of them by name.
- why not?Not because you're small. Because AI can't read your site, can't verify your business exists, and can't find you in the sources it quotes. Winners of AI answers had as few as 1–41 visitors from Google a month.
- what do I do?Three moves: make the site machine-readable, create your entity records, get into cited sources. With a market average of 35/100, the answer slots are still cheap.
three terms. thirty seconds.
GEO (Generative Engine Optimization) is the work of making AI engines — ChatGPT, Claude, Gemini, Perplexity — know your brand, trust it, and name it when a buyer asks "who should I pick?". SEO fights for positions on a results page. GEO fights for a place inside the answer itself.
GEO health score is a 0–100 rating of how ready a site is to be read, verified and cited by AI — technical readability, schema, entity records, trust signals and content format, rolled into one number. Below ~35 a brand is effectively invisible to the models; the engineering-heavy fixes usually add 15–25 points each.
ETV (Estimated Traffic Value) is what a site's organic traffic would cost if you bought the same clicks as ads — a dollars-per-month proxy for how much search visibility is actually worth. A brand with an ETV of $62/mo next to a competitor at $3,342/mo isn't losing a little traffic; it's donating the category.
same test. run 100 times.
Every one of the 100 audits runs the same reproducible protocol:
| parameter | value |
|---|---|
| period | april – july 2026 |
| AI engines | ChatGPT, Claude, Gemini, Perplexity, Grok, DeepSeek |
| prompts per brand | 2–14 live category prompts ("who should we pick?") + branded probes ("tell me about X", "is X legit") |
| languages | 1–5 languages of the target market (EN, ES, PT, RU, UA) |
| crawler checks | curl as GPTBot / ClaudeBot / PerplexityBot — HTTP status + actual bytes of text served |
| entity layer | Wikidata, Wikipedia, Crunchbase, Google Knowledge Graph, Common Crawl, fact consistency across profiles |
| machine readability | JSON-LD schema (validated), llms.txt, robots.txt, sitemap, canonicals, no-JS rendering |
| trust (E-E-A-T) | authorship, /about and /team, sourced claims, number consistency, external reputation |
Each audit ends in a GEO health score from 0 to 100. Every case below traces back to a specific audit — brands are anonymized under NDA.
100 audits, 5 mistakes on repeat.
These aren't occasional findings. They're systemic — present in nearly every audit regardless of niche, size or budget.
| # | mistake | frequency | what it kills |
|---|---|---|---|
| 1 | no entity anchor (Wikipedia / Wikidata / Crunchbase) | ~100% | AI can't verify you exist |
| 2 | no JSON-LD schema markup | 94% | AI can't parse who you are or what you sell |
| 3 | anonymous content, E-E-A-T vacuum | 89% | AI doesn't trust or cite you |
| 4 | no llms.txt | 87% | AI crawlers don't know what to read |
| 5 | site unreadable to AI crawlers (JS / 403 / weight) | 50% | AI physically cannot see your content |
to AI, you don't exist.
Not a single audited brand had a Wikidata record. Almost none had Wikipedia or a consistent Crunchbase.
Wikipedia and Wikidata are the top-trust sources for LLMs: models train on them, verify entities against them, and Google's Knowledge Graph is built from them. When a user asks "who is X and can I trust them", the model first looks for independent proof the brand is real. No record — no proof — no recommendation. Add the record and your entity score, health score and odds of entering the answer all climb.
| case from the audits | consequence |
|---|---|
| financial company with $2B processed volume — zero entity records | Claude replies "I cannot identify this firm" — and recommends competitors |
| startup whose Crunchbase describes last year's product | AI retells the outdated story as fact — the brand feeds its own misinformation |
| brand with three historical positionings across profiles | AI blends all three and hallucinates features that don't exist |
The fix: a Wikidata item, one canonical fact set across Crunchbase/LinkedIn, and — for mature brands — Wikipedia via Articles for Creation. The cheapest health-score points in all of GEO.
The standard picks are Chainalysis, Elliptic and Sumsub for KYT and wallet screening — all well-documented and widely used by exchanges.
"████████" doesn't come up in my reliable sources — I can't verify its certifications or client base.
a business with no machine-readable passport.
Typical schema score: 0–4 out of 100. Zero JSON-LD — no Organization, no Product, no FAQ, no Person.
JSON-LD is how you tell the machine directly: "we are company X, the product is Y, the price is Z, the license is here." Without it, AI has to guess from prose — and guesses badly, or not at all. The brands winning AI Overviews in our benchmarks win them substantially on technical markup — reproducible with pure engineering work, without a single article.
We also saw schema that "exists" but is broken: FAQ markup failing validation, a generic Organization with no business type, sameAs pointing to messengers or someone else's account. All of it costs machine trust.
The fix: the full stack — Organization + Product/Service + FAQPage + Person + BreadcrumbList. Typically one of the two largest single lifts to a health score (+15–25 points).
https://████████.com
✗ 0 items detected
an anonymous business has nothing to be trusted for.
No /about or /team (or they 404), articles with no authors, numbers contradicting each other across pages, a different HQ in every social profile.
LLMs weigh content through E-E-A-T logic: experience, expertise, authority, trust — hardest of all in YMYL niches (finance, gambling, legal). Anonymous text weighs less than the same text with a name, credentials and Person schema. De-anonymization is the rare fix that lifts your entire content corpus at once.
| case from the audits | the absurdity |
|---|---|
| law firm: 1,003 of 1,004 articles anonymous | the founder wrote a national digital-assets law — AI never names the firm, 0 of 5 answers |
| education project: 896 expert articles under pseudonyms | content depth scores 75/100, anonymity drags the corpus to 52 |
| product: cashback "2%" on one page, "5%" on another | the most-cited LLM source about the brand prints the wrong number |
| B2B vendor: template logos in the "trusted by" strip | AI and journalists see fake references |
The fix: live /about and /team, real authors with Person schema on every article, one canonical fact set everywhere. Cheap, fast — and the foundation everything else stands on.
by ████ team · no date · no author page
llms.txt doesn't exist.
Missing or returning 404 in 87% of audits; where it existed, it was often a link dump of the blog.
llms.txt is a sitemap written specifically for AI crawlers: who you are, what's worth reading, where the proof lives. The standard is young — and that's the point: your competitors don't have one either. In our competitive benchmarks, not a single direct rival of an audited brand had deployed it. The cheapest first move in GEO: an immediate signal to every AI crawler — and a head start for as long as the market keeps sleeping.
The fix: an llms.txt describing your entity, products and canonical URLs — updated on every release. Plus an explicit allow for GPTBot / ClaudeBot / PerplexityBot in robots.txt.
$ curl -I https://████████.com/llms.txt ← the audited brand HTTP/2 404 $ curl -I https://████████.io/llms.txt ← market leader HTTP/2 404 $ curl -I https://████████.xyz/llms.txt ← closest rival HTTP/2 404 → the whole category returns 404. the slot is unclaimed.
AI physically cannot read the site.
The most underrated mistake, because from the outside everything looks fine.
AI crawlers are not Googlebot. GPTBot, ClaudeBot and PerplexityBot do not execute JavaScript. A client-side-rendered SPA serves them an empty shell. An aggressive WAF serves them 403. Google may see everything — so Search Console shows an illusion of health while the models see nothing.
| case from the audits | google sees | AI sees |
|---|---|---|
| site with Lighthouse SEO 100/100 | full content | 1,133 bytes of empty shell |
| casino with brand demand up +15,900% YoY | healthy brand SERP | 403 on every request — the AI's answer is assembled from ScamAdviser (0/100) and complaints |
| fast site: Performance 0.95, LCP 2.4s | indexed pages | 1 byte of text, zero H1, a title of three dots |
| platform with billions in volume | a normal index | 3 URLs in Common Crawl vs 500+ for competitors — models cannot cite it |
| fintech service | a normal site | double blackout: CSR + robots.txt explicitly blocking AI bots |
The fix: SSR/prerender for key pages, a WAF audit against AI crawlers (one curl to check), a weight diet for heavy pages. Often a single, well-scoped engineering task — and the only thing between the brand and the models' corpus.
$ curl -A "GPTBot/1.0" https://████████.xyz/ <!doctype html><html><head> <script src="/assets/app-4f2c.js"></script> </head><body> <div id="root"></div> </body></html> → 1,133 bytes · 0 words of content · 0 links lighthouse seo score of this page: 100/100
Not sure how many of these five your business has? We'll check — six engines, your category, your health score.
request a manual audit →87 of 100 audited brands: AI knows exactly who they are — but recommends someone else.
5 beliefs. 100 audits. 0 survivors.
The single most important finding of 100 audits: there is no correlation between organic traffic and entering an LLM answer. None.
| benchmark example | seo performance | llm visibility |
|---|---|---|
| niche B2B vendor | 29 visits/mo, 6 keywords | recommended by all 3 engines |
| its traffic-leader competitor | 8,336 visits/mo, ~3,000 keywords | holds no recommendation slot |
| event agency | 1 visit/mo | named by AI in 2 verticals |
| the biggest brand in its category | 270,000+ visits/mo | only #13 for its own head query |
| infrastructure company | DA 1–5, ~2.5 visits/mo | #1 pick in ChatGPT and Claude |
Why: an LLM builds its answer not from rankings but from sources it trusts — "best-of" listicles, industry directories, Wikipedia, readable whitepapers. Placement decides, not positions. Bad news for anyone who spent five years buying only classic SEO. Great news for everyone else: entry into the AI answer is cheap today and doesn't require years of domain authority.
In our data, categories were regularly won by businesses whose sites got 1 to 41 visitors from Google a month. Their shared trick is page format: a dedicated page per vertical or use case instead of one generic "services" page, a comparison page instead of a product page. One structurally correct page beat competitors' year-long content plans.
The gap we found in 87% of audits: branded recall ~93% — AI knows and accurately describes the brand. Category recall — 4.5% of answers across 66+ live prompts. AI knows exactly who you are, and in 95.5% of cases recommends someone else. In one audit the category slot was held by a "competitor" whose website didn't even resolve — its name simply lived in the sources the model cites. Knowledge ≠ recommendation. The recommendation must be built separately.
Lighthouse measures a human's experience in a browser with JavaScript executed. AI crawlers don't execute JS. That's how "Lighthouse SEO 100/100" and "the bot receives 1,133 bytes of nothing" coexist on the same page — we watched it happen. The only honest test is a curl as GPTBot. Takes a minute. The results almost always hurt.
| case from the audits | volume | result |
|---|---|---|
| site with 1,422 URLs | 1,000+ articles | 0 top-3 positions; a competitor 6.8× smaller gets 6× the traffic |
| blog with 136 posts | 136 URLs | a canonical bug collapsed all of them into 1 indexed URL |
| 896 articles of ~4,500 words | ~4M words | stuck on page 2: every author is anonymous |
Volume without structure, markup and authorship isn't an asset — it's ballast. The winner isn't whoever writes more; it's whoever the machine can read, verify and quote.
one disease. three faces.
One diagnosis looks different in every niche. Here's what 100 audits showed across three segments.
strong products. invisible brands.
The most paradoxical segment: strong products, real volume — and systemic invisibility.
| metric | typical value across audits |
|---|---|
| average GEO health score | 37/100 |
| branded recall | ~100% — AI knows nearly every brand |
| category recall | ~0 — "best crypto card / DEX / wallet" answers repeat the same 3–4 rival names |
| Wikidata / Wikipedia | 0 of audited brands |
| llms.txt | present at ~20% — the best score of any niche, still a failure |
| typical organic traffic | 10–200 visits/mo — against products doing millions in volume |
- The product-vs-visibility gap is the widest on the market. We audited products with billions in volume whose visibility was single-digit visits and zero AI citations. Web3 builds product before distribution — and AI punishes it mercilessly.
- Every category answer is already held by 3–4 brands — and they're not always the best products. They're the ones sitting in the listicles, with machine-readable whitepapers.
- Name collisions are the chronic disease. Short "cool" names collide with words in other languages, other products and meme brands. In one audit, 99% of a brand's "traffic" turned out to belong to someone else's similarly-named product.
- Categories are still cheap — but repricing fast. In several audits the category's LLM-visibility "leader" collected under 1,000 visits a month. One audited category grew 16.8× in 24 months; entry into AI answers by 2027 is priced at 2–4× today's cost.
Why it matters: if you're a Web3 brand, your category inside AI answers is either already taken by three names or still free. Either way the entry price right now is minimal: entity anchor + schema + readability + 2–3 placements in cited sources. In a year, the same entry will cost you a full-scale SEO war.
here AI doesn't ignore you — it warns people about you.
The harshest segment: here AI doesn't just stay silent about brands — it warns users against them.
| metric | typical value across audits |
|---|---|
| average GEO health score | 24/100 — the market's worst |
| category recall | 0 of 10 live prompts in the segment |
| negative AI tone ("high-risk", "possible scam") | 2 of 3 audited brands |
| schema / JSON-LD | 0/100 across all audited |
| AI crawlers blocked (403 / CSR) | 2 of 3 |
- The AI's verdict is assembled from your worst sources. When the site is closed to AI crawlers, the model builds its answer from whatever it can read: ScamAdviser, Trustpilot, forum complaints. We watched Perplexity quote a single unresolved $300 complaint verbatim — as the verdict on a brand. The license was real; it just wasn't machine-readable, and the complaint was.
- WAF self-sabotage is the niche's signature mistake. Cloudflare 403s every AI bot while letting Googlebot through. Result: Google looks great, AI calls the brand a possible scam — and the owner never sees it, because they only watch Search Console.
- Mirrors and clones destroy the entity. A brand smeared across 6–11 mirror domains reads to a machine like a phishing network — with everything that implies for trust.
- Affiliates own the SERP — and AI learned from them. Even a giant with 270K+ monthly visits ranks only #13 for its own head term: reviewers and Reddit own both the SERP and the LLM sources. The only content about you AI will read is what others wrote — or what you finally make readable yourself.
high-risk
Exercise caution: ██████ shows several red flags. ScamAdviser assigns the domain a trust score of 0/100, and there is an unresolved player complaint about a $300 withdrawal on AskGamblers. Independent confirmation of its licensing could not be verified. Consider established alternatives instead.
sources
Why it matters: in iGaming, GEO is first of all reputation defense. While your site is closed to AI, ScamAdviser and angry players write your brand's verdict. Opening the crawlers, marking up the license, building a trust hub isn't "marketing" — it's stopping a daily loss of deposits on the query "is it even legit?".
world-class portfolios the machine can't read.
The segment where hard proof of work — cases, awards, clients — systematically fails to translate into machine language.
| metric | typical value across audits |
|---|---|
| GEO health score | 28–45/100 |
| category recall | 0 of 13 live prompts ("who builds our booth", "best licensing law firm", "who to hire for development") |
| rival brands named in those same answers | ~70+ |
| anonymous content | up to 1,003 of 1,004 articles |
| typical organic | 3 branded keywords — or 190 keywords with zero top-3 |
- Hard proof in an unreadable form. International awards, NPS 87%, global-brand clients, 500+ projects — all of it lives in PDFs, images and portfolio pages, none of it in schema, entity records or cited sources. AI praises the agency by name — and never names it in a single "who should we hire" answer.
- One "services" page instead of a page per vertical — the mistake that costs the whole category. The winners in AI answers are agencies with a dedicated page per vertical or scenario. In our benchmarks an agency with 1 visit a month won 2 verticals on format alone.
- Expertise exists — authorship doesn't. Founders write laws and keynote industry stages, while 100% of the site's articles are anonymous. The model can't link a person's expertise to the company's content.
- The inversion happens too: we audited a company that was already the #1 pick of ChatGPT and Claude in its category — at DA 1–5 and near-zero organic. Its LLM visibility came from deep niche content. Without an entity foundation the position is fragile — but it proves the B2B slot is taken by content and entity, not domain authority.
Why it matters: a 2026 B2B buyer starts the shortlist by asking AI. If the answer names ~70 firms and never yours, you're losing tenders you never even heard about. The good news: services slots are the fastest to claim — the right vertical page + named experts + 2–3 listicles are what produce the first mentions.
read. verify. quote. that's the whole game.
A hundred audits reduce to one sentence: AI recommends whoever it can read, verify and quote.
| layer | what's in it | effect |
|---|---|---|
| readability | SSR/prerender, unblocking AI crawlers, llms.txt, page weight | AI sees your content for the first time |
| entity | Wikidata, consistent Crunchbase/LinkedIn, full JSON-LD stack, named authors | AI can verify you're real |
| citability | answer-shaped pages, listicles, industry sources, trust hub | AI starts naming you in the category |
The market's average health score is 35/100. That means almost any category is currently winnable from a standing start — by whoever makes the first three moves before the neighbor does.
asked every single time.
what is GEO and how is it different from SEO?
GEO (Generative Engine Optimization) is optimization for AI engines: ChatGPT, Claude, Gemini, Perplexity. SEO fights for Google rankings; GEO fights for your name inside the AI's answer to "who should I pick?". Our 100 audits show no correlation between the two.
how do I check whether AI can see my site?
Request your homepage with the GPTBot user-agent (a single curl command) and look at how much text comes back. An empty shell or a 403 means AI crawlers can't see your content — whatever Lighthouse says.
how long until we show up in AI answers?
It depends on where you start — which is exactly what the audit measures. The order is always the same: readability first, then the entity layer, then citations; the earliest movers on the health score are llms.txt, Wikidata and schema. We re-audit on a fixed cadence and track the lift against the baseline.
why does AI call my brand "risky" when we're licensed?
Because the license exists in a form machines can't read, while negative sources (ScamAdviser, complaints) are perfectly readable. AI rules on what it can read. Fixable: machine-readable license markup, a trust hub, opening AI crawlers.
which niches are the 100 audits from?
Web3 and crypto (cards, DeFi, exchanges, infrastructure), iGaming (casino, betting, B2B vendors), agencies and B2B services (development, events, legal). One methodology for all — see the top of this page.
does GEO replace SEO?
No — it runs alongside it. Google still matters, and many sources LLMs cite are themselves ranked pages. But the two don't correlate: our audits found brands with excellent SEO and zero AI presence, and vice versa. Treat them as two channels with separate scoreboards.
we block AI bots on purpose. is that a problem?
The biggest one we see. Blocking GPTBot or ClaudeBot doesn't remove your brand from AI answers — it removes your version of the story. The model still answers questions about you, using whatever it can read: review aggregators, complaint threads, competitors. In our audits, blocked sites consistently got the harshest verdicts.
can we just buy placements and reviews, like old-school SEO?
Careful. Models weigh independence: in one audit, the only readable corpus about a brand was its own paid promo — and the engines flagged the sources as "promotional, not independent" and issued a high-risk verdict. Bought volume without independent confirmation can make things worse, not better.
which AI engines actually matter?
We test six: ChatGPT, Claude, Gemini, Perplexity, Grok, DeepSeek. Weight them by where your buyers ask: Perplexity and Gemini pull heavily from the live web and citations, ChatGPT and Claude lean on training data and entity knowledge. The mistakes in this article affect all six — that's why they're the priority.
what exactly is in the manual audit?
The same protocol as these 100: live category and branded prompts across six engines, crawler checks with real AI user-agents, the full entity and schema review, and competitor benchmarks. You get a health score, the verdicts AI currently gives about you, and your three most expensive gaps.
your buyers ask AI first. right now it's recommending someone else.
request a manual GEO audit
and see which queries you're missing, who's winning them, and why.