Every claim we make, graded against the sources

What the evidence saysRoughly half of the domains cited in Google AI Overviews do not appear in the organic top 10 for the same query, and URL-level overlap between Google organic, AI Overviews and Gemini is very low (Jaccard ≈ 0.11–0.18). The citable surface genuinely diverges from the ranked surface.

So whatThis is the strongest empirical justification for measuring AI visibility as its own surface. Your rank report does not tell you whether you are cited.

What the evidence saysRepeat the same prompt and you get materially different sources: day-to-day source overlap runs ≈ 0.34–0.42. In one analysis 57.8% of ChatGPT runs never triggered a web search at all — the model answered from parametric memory, citing nothing.

So whatAnyone who shows you a single screenshot of ChatGPT naming (or omitting) your brand is showing you noise. Only repeated sampling over a window produces a number you can act on. This is why we sample every prompt multiple times across a 14-day window.

What the evidence saysEach major engine runs multiple separately-blockable agents that split training, live retrieval and search indexing. This is documented in OpenAI's, Anthropic's, Perplexity's and Google's own crawler docs. A blanket 'block AI bots' rule is a common misconfiguration that silently removes a site from the citable surface.

So whatThe least glamorous item on this page is the one most likely to be costing you citations today. It is free to fix.

What the evidence saysThis originates with the GEO paper (Aggarwal et al., KDD 2024), which reported gains up to ~40% for tactics like Quotation Addition and Statistics Addition. But C-SEO Bench (NeurIPS 2025) re-tested seven of those same methods and found most are 'largely ineffective' and 'frequently have a negative impact on document ranking' — while traditional retrieval/ranking work was significantly more effective. It also found the gains are congested and zero-sum: as more competitors adopt the same tactic, the advantage decays toward zero.

So whatContent rewriting is not a reliable lever on its own, and it is not the moat vendors imply. Getting retrieved at all dominates what you do once retrieved. We now treat extractability as hygiene, not strategy.

What the evidence saysThe figure is real but heavily qualified. It is a redistributive share metric across a fixed set of five sources that sums to one — not an absolute lift, and explicitly 'does not mean that 40% more readers will click'. It was measured on GPT-3.5-turbo in a custom two-step harness, never on production ChatGPT/Gemini/AI Overviews, and is conditional on the source already sitting in the context window. It has no 2026 replication. The paper's own abstract concedes efficacy 'varies across domains'.

So whatCite the paper for vocabulary — it coined the term GEO. Do not cite the 40% as a promise. Any agency quoting it unqualified either hasn't read it or is hoping you haven't.

What the evidence saysGoogle's official position is that it is not: generative AI features are rooted in core Search ranking and quality systems, so 'optimizing for generative AI search is optimizing for the search experience, and thus still SEO.' The counterweight is the citation-divergence evidence above — AI engines demonstrably cite a different source set, and ChatGPT/Claude/Perplexity do not run on Google's ranking at all.

So whatOur honest read: for Google's own surfaces, Google is largely right — GEO is mostly good SEO plus crawler hygiene. For the non-Google engines, it is a genuinely separate surface that your rank tracker cannot see. That's the part worth paying for, and we'd rather say so than pretend the whole thing is new.

What the evidence saysIt is a vendor coinage, including ours. There is no authoritative or standardised definition, no agreed formula, and no cross-vendor comparability — two vendors' 'AI Share of Voice' numbers are generally not the same measurement.

So whatOurs is defined openly in our methodology so you can audit it. But treat any vendor's SOV number as internal to that vendor, and never compare two vendors' figures directly.

What the evidence saysNo major engine documents reading llms.txt from third-party sites. Google explicitly names llms.txt and states it does not use it. OpenAI's, Anthropic's and Perplexity's crawler docs cover robots.txt, user agents and IP ranges, and never mention consuming llms.txt. Vendor claims that Perplexity 'supports' llms.txt trace to uncited marketing. (Several of these companies publish an llms.txt for their own docs — publishing is not consuming.)

So whatShip llms.txt if you want — it's an hour and it's harmless. Do not buy it as a visibility lever. We ship one and claim nothing for it.

What the evidence saysGoogle fully removed FAQ rich results on 2026-05-07. FAQPage remains a valid schema.org type and Google states unused structured data does not harm Search — but it confers no visible Search benefit today. Whether FAQ markup affects citation in non-Google engines is an open question with no good public evidence.

So whatWe still ship FAQ blocks — because clear question-and-answer prose is genuinely easy for a model to lift, regardless of markup. We do not claim the schema itself earns you anything in Google.

What the evidence saysWe looked for primary evidence on time-to-first-citation and found none. There is no public study establishing how long AI-visibility work takes to show effect. The 60–120 day figure circulating in this industry — including, previously, on this site — is a vendor convention, not a finding.

So whatWe removed it as a promise. What we can honestly say: technical fixes (crawler access, indexing) can land within days, and citation change is observable in weekly sampling. We will not put a number on your results that we cannot source.