Methodology v2026.11
How the audit actually works
Most AI-visibility audits are one prompt to one model, dressed up as analysis. This page explains what ours measures, how it's scored, what we refuse to claim — and a dated changelog of every methodology change, so you can see the audit is maintained, not fossilised.
Maintained by James Eltherington, founder.
Measured AI visibility
We don't infer your AI visibility from on-page signals — we measure it. The audit builds a set of questions your customers actually ask, puts them to ChatGPT, Perplexity, Gemini, Google AI Overviews, Grok and Claude, and records the answers: whether you appear, where you rank, and who gets recommended instead of you. The queries and results are in your report — you can re-run any of them yourself and check.
Engines don't give the same answer twice, so a single query proves nothing — that's why this is a structured probe across engines and queries, scored in aggregate, rather than one screenshot of one lucky answer.
How we read your site
We read the HTML your server returns and do not run JavaScript, because that is what the AI crawlers do: GPTBot, ClaudeBot and PerplexityBot all fetch the raw page. Content that only appears after JavaScript runs will not appear in your audit, and it does not appear to them either. When a site is built that way we say so at the top of the readiness report rather than let a low score look like a mistake. The one exception is a bot wall: when a site answers with a challenge page we render the homepage in a real browser to get past it, and the report still records that the wall is there.
How scoring works
Your Readiness score has two equal halves. One is deterministic: measured signals from your live site, evaluated in code against a fixed rubric. The other is an AI analysis of the same evidence, which catches what code cannot, such as weak content or a confusing brand name; its findings are checked against the measured signals. Because of that second half, an unchanged site can move a few points between runs, so read a small move as noise rather than news.
Your AI Visibility score comes from the measured probe above. The two blend into one Overall score across six categories:
- AI citability — Whether your content is structured so an AI engine can lift a passage and cite it — self-contained answers, extractable facts, question-shaped structure.
- Brand authority — Whether your brand exists in the places AI engines actually draw citations from — community discussion, reference sources, video, professional networks.
- Content quality (E-E-A-T) — Experience, expertise, authoritativeness and trust signals — the framework engines use to decide whether a source deserves to be cited at all.
- Technical foundations — Whether AI crawlers can physically reach and read your content: crawler access, server-side rendering, site health. AI crawlers do not run JavaScript — sites fail here silently.
- Structured data — Machine-readable entity information — who you are, what you offer — validated, not just detected.
- Platform readiness — Per-platform checks for Google AI Overviews, ChatGPT, Perplexity, Gemini and Copilot, because each selects sources differently.
We publish the categories, not the individual checks and weights — that's the part competitors would like us to hand over.
Sentiment: what the engines say about you
Your Sentiment score is the tone of the answers that name you, from 0 to 100, where 50 is evenly balanced. It needs at least five mentions to be a score at all, and it is reported on its own: it is never blended into Overall, because being spoken of warmly is not the same as being found. Paid runs add two fixed reputation questions to your buyer questions, and each report says how many mentions came from them. A free Check does not include sentiment.
Alongside the tone, the report shows what the engines credit you for, their exact words, and the facts they state about you, checked against your own site. It also shows what each buyer answer that named you did with you: picked you alone, shortlisted you, listed you, mentioned you in passing or steered the buyer away. Tone says how they spoke of you; this says whether they chose you.
What a good visibility score looks like
A score out of 100 tells you nothing on its own, so we publish what we have actually measured rather than leaving you to guess. These are every run on record as at 5 November 2026, one run per domain, the most recent, paid runs only. A free Check asks each engine once on four engines and is not the same measurement as a plan run, so it is not counted here.
Household names: 52 to 79
We run sites the engines could hardly fail to know, as positive controls. stripe.com scores 79, wikipedia.org scores 74, bbc.co.uk scores 65. Nothing we have measured scores higher than 79, because no brand is named in every answer to every buyer question. That is the practical ceiling, and it is a long way below 100.
Real businesses: 0 to 73
Across the 18 customer and prospect sites we have measured, the middle score is 11. 11 of them score under 20 and 4 have never been named in a single answer by a single engine. A score in the teens is the ordinary starting point for a healthy business, not a failure.
Every report carries this comparison under its own score, so you always know whether your number is good. We would rather tell you that 11 is normal than let a number out of 100 imply a crisis and sell you something on the strength of it.
The QA gate & refund
Before a report reaches you it passes an automated QA gate: completeness of every section, presence and consistency of scores, and evidence backing the findings. A report that fails the gate is automatically retried; if it still can't pass, it is held for human review. If we still can't deliver, a one-off purchase is refunded in full and a run paid with credits gets its credits back. You will never receive a report that didn't pass.
What we refuse to do
The GEO industry has a snake-oil problem. These are written rules enforced in our report pipeline — not marketing copy:
No projected traffic or revenue figures.
“This fix is worth £4,000/month” sounds great and is unmeasurable. Our reports may only state impact we observed — e.g. “ChatGPT's crawler cannot retrieve your pricing page, so ChatGPT cannot cite it.”
No guarantees of rankings or citations.
AI answers are probabilistic — anyone guaranteeing you a citation is guessing. We measure where you stand and fix what's measurably in your way.
No stats without a source.
Every research claim in our reports traces to published research or our own measurements. When evidence is correlational, we say so.
No selling fixes the evidence doesn't support.
When the evidence for a tactic weakens — as it has for some once-fashionable GEO tactics — we downgrade or drop it, and note it in the changelog below.
We run it on ourselves: current score 42/100
We run the same pipeline on citeify.ai, with the same scoring and no special treatment. Our latest published result, from 14 October 2026 under v2026.10.3: Readiness 84/100, measured AI Visibility 0/100, Overall 42/100. Zero visibility means that when 10 buyer questions in our own niche were each put five times to six engines, none of them named or cited citeify.ai. We publish that willingly: the site is ready to be cited and the brand is not yet known, which is the honest starting point of a new business and the number every future re-audit gets measured against.
The instrument is on display too. Across our last five Audits the site's Readiness has read 84, 81, 84, 81 and 84. Some of that is the site and the methodology changing and some is the AI half of the score, and we would rather show you than smooth it over.
Methodology changelog
AI search changes monthly; an audit that doesn't is worthless. Every change to what we measure is versioned and dated here. Your PDF report carries the methodology version it was produced under.
v2026.11
November 2026
- Recorded late, because this changelog should have carried it when it shipped: since September 2026 paid runs have also measured sentiment. That is a tone score from 0 to 100 for the answers that name you, the things the engines credit you for, their exact words, two fixed reputation questions, and the facts they state about you, checked against your own site. Sentiment is reported on its own and is never part of Overall.
- Also recorded late: since October 2026 reports have labelled what each buyer answer that named you did with you, from picking you to steering the buyer away. In November 2026 the rules were tightened after a hand check found them over-claiming. "Recommended" now means picked alone; a joint pick, or a pick for one part of the need beside others, counts as shortlisted, and an entry in a list or table counts as listed. The labelling also moved to a stronger model, Claude Sonnet 5.5. The chart shows only on runs labelled under the new rules.
- Corrected how we describe Readiness. This page said the same site produces the same score every run. Half of the score is an AI analysis of the measured evidence, so an unchanged site can move a few points between runs. How the score is calculated has not changed; what we said about it was wrong.
- The figures for what a good visibility score looks like are re-measured: 18 real sites, with a middle score of 11 and a highest of 73.
v2026.10.3
October 2026 (update 3)
- The audit's advice for Google AI Overviews now follows what Google publishes: there are no special requirements, and no llms.txt, AI text file or special structured data is needed to appear there. Readiness for AI Overviews is judged on whether Google may quote your pages, how well they rank and how clearly they answer, and it no longer credits schema markup.
- llms.txt is treated as optional: a missing one is at most a low-priority finding, and it is no longer one of the essentials on your checklist. Structured data is still assessed, because ChatGPT, Perplexity and Gemini use it to recognise who a business is.
v2026.10.2
October 2026 (update 2)
- The audit now reads Google's snippet controls on every page it crawls: nosnippet, max-snippet and noindex in the robots meta tags and the X-Robots-Tag header, and data-nosnippet on page elements. Google keeps a page that uses nosnippet or max-snippet:0 out of AI Overviews and AI Mode, so a site that sets them on pages it wants quoted now gets a critical finding, and the essentials checklist shows whether Google may quote you.
- Google-Extended is now treated as what Google says it is: a robots.txt setting that controls whether your content is used for Gemini, not a crawler, and with no effect on Google Search, AI Overviews or AI Mode. Reports measure it separately and no longer count opting out with it as blocking AI crawlers. A site that opted out with Google-Extended alone was marked down on Platform as if it blocked AI crawlers, and scores higher on Platform from this version.
v2026.10.1
October 2026 (update 1)
- The written summaries that open the AI visibility, readiness and fix sections of a report are now drafted by Claude Haiku 5.5 instead of Haiku 4.5. They are still written only from the figures we measured and still pass the same checks before a report is released. Nothing that is measured or scored has changed: every score, verdict and tone figure is produced exactly as before.
- Reports now say when an engine cited your site as a source but never named your business in its answer.
v2026.09.11
September 2026 (update 11)
- The ChatGPT engine now asks the model ChatGPT gives its free users, GPT-5.6 Luna, which OpenAI made the default in August 2026. Until 22 September it asked GPT-4o, which OpenAI has since retired from ChatGPT, so the answers we measured were from a model buyers were no longer talking to. ChatGPT verdicts from before this date were measured on the older model; the other five engines are unchanged, and no stored score has been altered.
- Every engine call is now priced from the provider's own usage figures rather than a per-call estimate, so the cost figures we hold are what we are billed. This changes nothing in any report.
v2026.09.10
September 2026 (update 10)
- The AI visibility score now rests on two things only: how often the engines name you, and how often they cite your site as the source. Share of voice, which measured how many other businesses were named alongside you, has been removed from the score. It rewarded appearing in answers that mention few competitors, which are rarely the answers buyers ask, and it marked down brands that show up in crowded comparisons, which are. Share of voice is still measured and still shown in your report; it is no longer part of the headline number.
- Every score on record has been recalculated on the new basis, so your trend line stays comparable end to end and no run is measured one way and the next another. Scores that leaned on share of voice have fallen and scores built on being named and cited have risen; the date each run was measured has not changed.
- We now publish what a good score is. Household names we run as controls score in the fifties and sixties, no site we have measured has scored above 70, and the middle score across the real businesses we have measured is in the teens. Every report says where its number sits among them, and /methodology lists the figures.
v2026.09.9
September 2026 (update 9)
- Full audits and Tracker runs now put every question to six engines: Grok and Claude join ChatGPT, Perplexity, Gemini and Google AI Overviews. The free check keeps the original four. The visibility score averages whichever engines answered, so scores from before this update remain comparable engine by engine.
- The AI visibility score is shown as a band: the same run recomputed on each of its five samples, so a two-point wobble is not read as a move. When the question set has been edited, the change since the previous run is computed on the questions both runs asked.
v2026.09.8
September 2026 (update 8)
- You can now steer the buyer questions we ask the AI engines. Tell us what matters in a few lines and we rewrite the set from it, or edit, pin, remove and add the exact questions yourself, on the dashboard for each domain. Up to ten are included; more can be added.
- A first paid audit now waits for you to check its questions before it runs, and starts on its own after a day if you do nothing.
- Changing the questions starts a new baseline: the trend line marks the run where the set changed, and each question in the table shows its own verdicts from earlier runs, so unchanged questions keep their history and edited ones start afresh.
v2026.09.7
September 2026 (update 7)
- The PDF report is now rendered directly from the measured data, with no model in the loop. It carries the same content as the dashboard in the same order (overview, AI visibility, readiness, the fix), so the two can never disagree, and any past report can be re-issued in seconds under the current design.
- The executive summary and quick wins written during the audit are now kept with the audit and printed in the PDF. Reports produced before this update are re-issued with a summary built from their scores.
- The AI visibility and readiness sections each open with a short written summary of what was measured: which questions were lost and to whom, and which categories and findings matter most. Shown on the dashboard and in the PDF.
v2026.09.6
September 2026 (update 6)
- The report is reordered around the question you actually asked. It now opens with the verdicts: every buyer question, a cited / mentioned / absent result per engine, and the competitors recommended instead. The readiness diagnosis follows as "why", and the findings, 30-day plan and coding-LLM briefs close as "the fix". Each report has its own page with deep-linkable sections.
- Every question now maps to your own site: when an engine cited you, the report names the page it cited; when you were absent, it names the crawled page best placed to answer that question, or says plainly that nothing on the site does yet.
- A new panel puts "built to be cited" (your citability readiness) next to "actually cited" (the measured citation rate across engines), with a one-line reading of the gap.
- Full audits now ask each question five times per engine instead of three, so a verdict reflects a majority of samples rather than a lucky or unlucky answer.
- Free checks now read sites behind bot-protection challenges with the same headless fallback the full audit uses.
v2026.09.5
September 2026 (update 5)
- The £19.99 Quick Scan is retired and replaced by a free Visibility Check. The old scan graded your homepage foundations and never asked an AI engine anything; the free check runs the same 10 buyer questions the full audit uses, on the same four engines, once each, and shows you the verdict per question and the competitors recommended instead. Your email is confirmed before the check runs.
- A free check and a later full audit of the same domain now ask the same questions, so the audit deepens the table you have already seen (three samples per question instead of one) rather than starting a new one.
- The "ready" email now leads with the headline from your report, such as how many questions you were absent from and who was recommended in your place.
v2026.09.4
September 2026 (update 4)
- A full audit no longer ships if its measured AI-visibility check could not run. Previously, if the answer engines were unreachable, the report would quietly go out without the visibility page and the headline score would fall back to readiness alone. Now the audit is held for human review and re-run, so every published full audit carries a real measurement.
- Corrected what we claim to measure: live AI visibility is measured on ChatGPT, Perplexity, Gemini and Google AI Overviews. Bing Copilot has no queryable answer service, so we score readiness for it but do not present it as measured. Removed references to features that are not yet available to buy.
v2026.09.3
September 2026 (update 3)
- The buyer-intent questions behind your visibility score are now tailored to what your business actually does and where it operates. A local or regional business is measured on the questions its real customers ask — including “near me” and place-specific searches — and against competitors at its own scale, not global market leaders it never competes with. This makes the visibility score, and the competitors it names, far more relevant and actionable.
v2026.09.2
September 2026 (update 2)
- Our crawler is now a well-behaved, clearly identified visitor: it spaces out its requests so it never trips a site’s rate-based security, and it announces who it is so hosts can recognise it.
- When a site’s security serves our crawler a “please wait, verifying…” challenge, we now read the real page by rendering it in a full browser — so good content behind a strict security wall is measured fairly instead of scored as empty. Crucially, the report still flags that the wall exists: AI engines like ChatGPT and Perplexity cannot pass these challenges, so a site behind one is effectively invisible to them, and we call that out as a critical issue to fix.
v2026.09.1
September 2026 (update 1)
- Strengthened the crawl-integrity safeguards after a real-world miss: bot-challenge detection now recognises many more challenge-page variants, checks the exact pages our crawler was served (catching intermittent challenges), and retries once before concluding a site cannot be read. Found by re-auditing a site we knew was healthy.
v2026.09
September 2026
- Added crawl-integrity safeguards: if a site serves our scanner a bot challenge instead of its real content, or a score moves implausibly far between audits, the audit is now held for human review rather than published. No one should ever receive a confident report scored against a security wall.
- Anything our crawler is blocked from reading is now always reported as "could not be verified", never as missing from your site.
- If an AI engine is unavailable while your audit runs, the report now says so explicitly; your visibility score is averaged only over the engines that actually answered.
v2026.08.4
August 2026 (update 4)
- Reports now show the actual buyer-intent questions we put to each AI engine, with the outcome per engine for every question (cited, mentioned, or absent) and which competitors were recommended instead. The measurement itself is unchanged: this data was always collected, and it is now published in full so you can verify any result yourself.
v2026.08.3
August 2026 (update 3)
- Audits now always include the site’s About/team page when judging authorship and expertise — a named expert who exists but isn’t surfaced is now reported as “surface it”, not “missing”. Found by re-auditing our own site.
- Hardened finding categorisation so every issue files under the category that owns it (a security finding can no longer appear under Content Quality).
- FAQ guidance sharpened: visible question-and-answer content is what earns citations; FAQ markup alone is no longer recommended as a fix following Google’s May 2026 rich-results retirement.
v2026.08.2
August 2026 (update 2)
- Modernised AI-crawler coverage to the current 2026 user-agent landscape — including the newer Anthropic, Mistral and DuckDuckGo crawlers — from a single canonical list.
- Added a CDN-level access check: some CDNs now block AI crawlers by default even when robots.txt looks open. The audit now tests how AI crawlers are actually served.
- Down-weighted llms.txt to an informational check, in line with the evidence that major engines don’t consume it (it retains minor value for documentation sites).
- Recalibrated structured-data guidance to current evidence: schema is scored for entity clarity and machine readability, not as a claimed citation booster; FAQ rich-result guidance updated for Google’s May 2026 deprecation.
- Corrected platform facts across the audit (Google-Extended vs AI Overviews controls, ChatGPT’s own search index, Microsoft Copilot naming) and added machine-readable freshness detection.
v2026.08
August 2026
- Unified the scoring model so every layer of the audit — analysis, report and PDF — is produced from a single documented rubric.
- Visibility probe now samples each question multiple times per engine (AI answers vary run to run — one sample is a coin flip) and re-uses a domain’s question set between audits, so score changes measure visibility movement, not question drift.
- Tightened our evidence policy: every claim in a report must trace to a measured signal or published research. Estimated traffic and revenue projections are now prohibited in our reports.
- Recalibrated research citations against the current academic literature on generative engine optimisation (2024–2026).
- Every report PDF now carries its methodology version, tied to this changelog.
v2026.07
July 2026
- Added the measured AI-visibility probe: your customers’ real questions are put to ChatGPT, Perplexity, Gemini and Google AI Overviews, and the answers recorded.
- Moved readiness scoring to deterministic, code-computed checks grounded in measured site signals — the same site now always produces the same score.
- Added multi-page citability analysis and automated QA gating on every report.
See it applied to your site.
About 25 minutes from domain to report. QA-gated, with your money or credits back if we cannot deliver.