One hundred audits, one protocol, six engines. Five failures came back so often that they stopped looking like mistakes and started looking like the default state of the market.
| # | 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 |
01 · to AI, you don't exist
Found in ~100% of audits. 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 language models: they train on them, verify entities against them, and Google's Knowledge Graph is built from them. When a buyer 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.
| case | 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| startup whose Crunchbase describes last year's product | AI retells the outdated story as fact — the brand feeds its own misinformation |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 and LinkedIn, and — for mature brands — Wikipedia via Articles for Creation. The cheapest health-score points in all of GEO.
from a live category probe, audit #41 — "which crypto AML screening tools should a startup exchange consider?" The model named three standard vendors, then said the audited brand "doesn't come up in my reliable sources — I can't verify its certifications or client base." That vendor is ISO-certified with hundreds of clients. None of it was machine-readable, so none of it existed for the model.
02 · a business with no machine-readable passport
Found in 94% of audits. 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 the model guesses from prose — badly, or not at all. The brands winning AI Overviews in our benchmarks win them substantially on technical markup, reproducible with pure engineering work and not a single new article.
We also found schema that "exists" but is broken: FAQ markup failing validation, a generic Organization with no business type, sameAs pointing at messengers or someone else's account. All of it costs machine trust.
The fix. The full stack — Organization, Product or Service, FAQPage, Person, BreadcrumbList. Typically one of the two largest single lifts to a health score: +15–25 points.
03 · an anonymous business has nothing to be trusted for
Found in 89% of audits. No /about or /team — or they 404. Articles with no authors. Numbers contradicting each other across pages. A different HQ in every social profile.
Models weigh content through E-E-A-T logic: experience, expertise, authority, trust. It is hardest of all in YMYL niches — finance, gambling, legal — where an anonymous page weighs less than the same page with a name, a credential and a face behind it.
04 · llms.txt doesn't exist
Found in 87% of audits. Missing or returning 404; where it existed, it was often a link dump of the blog.
It is the cheapest file in this list and the one most often skipped. It tells an AI crawler what on the site is worth reading and in what order — and it takes an afternoon.
05 · AI physically cannot read the site
Found in 50% of audits. Client-side rendering, 403s to AI user agents, page weight that times the crawler out. A full blackout in a third of them.
Lighthouse measures a human's experience in a browser with JavaScript executed. AI crawlers don't execute JavaScript. That is how a green Lighthouse score and an empty shell for bots live on the same page — we watched it happen.
the honest test takes a minute: curl your homepage as GPTBot and read what comes back. In one audit the answer was 1,133 bytes and zero words of content.
what the five have in common
None of them is about writing more. Every one is about being readable, verifiable and quotable — in that order. A hundred audits reduce to that sentence.