How we test — and why you can trust the verdict
This lab reviews GEO/AEO tools it doesn't sell. The method below is the same whether a review is free or part of a paid audit.
Every site is scored on llms.txt presence, AI-crawler access, answer-first structure, JSON-LD schema, and entity clarity. The free checker automates this; an audit re-runs each signal by hand and records the reason behind each pass, warn or fail.
AI visibility isn't a keyword rank — it's whether an engine names you inside its answer. We sample real buyer prompts across ChatGPT, Perplexity, Gemini and Google's AI Overviews and log who gets cited for each. Because LLMs sample from probability distributions, a single run is noise: we ask each prompt 3 times on the live-web engine and report the mention rate with a 95% confidence interval (Wilson). We publish our N; see the variance yourself with the free Answer Consistency Test.
For each citation won or lost, we trace it to the page and the signal that earned it — the answer capsule, the schema, the source authority — so a fix targets a cause, not a symptom.
AI engines and tools change weekly. Each finding carries the date it was checked and its source, so a stale result is never mistaken for a current one. Pages carry a 'last checked' date for the same reason.
The research base
The five signals aren't taste — each one is grounded in primary sources: Google patents, the DOJ v. Google trial record, the May 2024 Content Warehouse leak, and peer-reviewed GEO research. The distilled map, with what each source actually establishes:
Google's documentation is explicit: a page that isn't indexed can't appear in AI Overviews or AI Mode. The DOJ v. Google trial record describes the same architecture from the inside — retrieval from the index feeds a customized Gemini that writes the answer.
The 2024 Content Warehouse leak shows Google resolving full redirect chains and keying document attributes to the post-redirect URL (urlAfterRedirectsFp, forwardingdup) — which is why our probes follow and re-validate every hop.
Trial testimony describes AI Overviews' model working in tandem with Search's ranking systems and the Knowledge Graph — entity data is a separate input to the answer. A site whose entity doesn't consolidate can't be grounded as one.
The leak exposes authorReputationScore as a per-document attribute — author reputation tracked at entity level, not as a text string. That's why the audit checks Person/Organization markup and sameAs links, not just page-level schema.
Google's own patent on generative summaries describes an LLM composing the answer from a pool of retrieved passages via a two-part prompt. The unit of citation is the passage — a 40–60-word self-contained block either exists on your page or on your competitor's.
Google's information-gain patent scores content on what it adds beyond already-selected sources. Repeating the consensus earns nothing; per-block unique facts and numbers do. Related patents cover phrase-based indexing and contextual term weighting.
The Princeton GEO benchmark (KDD '24) quantified content edits against generative-engine visibility: adding statistics lifted visibility by up to ~41%, and adding source citations helped lower-ranked pages most. Our answer-first checks encode those levers.
Google's stateful-chat patent application and its AI-features documentation both describe query fan-out: the engine decomposes a question into parallel sub-queries, and your page can be cited for a sub-query it never targeted. Visibility must be measured on a basket of prompts, not one keyword.
Trial testimony describes FastSearch/RankEmbed — an embedding-based retrieval path feeding AI answers, distinct from the classic ranking stack. Ranking #1 organically does not imply appearing in the AI answer; we log the answers themselves.
PoisonedRAG showed that ~5 injected documents can steer a RAG answer with ~90% success — proof that the retrieved-source set, not the whole web, determines what the engine says. Citations are winnable at the margin; diversity-aware selection (GIST) rewards being the different source.
Provenance note: leak attributes and trial testimony describe Google's systems as of their documents' dates — we treat them as evidence of architecture, not as guarantees of current behavior. Every link above was fetched and title-verified on July 13, 2026.
We review checkers, rank trackers and GEO platforms we don't own. When a free option does the job, we say so. The one thing we sell is our own GEO audit, disclosed everywhere it appears — and our verdicts on third-party tools don't follow the commissions.