Methodology · v1.0.2 · 26 Sep 2026
How we measure AI visibility
This page defines every term and procedure we use, so a skeptic can check our work. It is versioned; changes are listed at the bottom.
1. Definitions
Every term links to its glossary entry on first use.
- Answer engine
A system that answers a question in prose, with or without sources, such as ChatGPT, Perplexity, Gemini, Google AI Mode and AI Overviews, Claude and Copilot.
- AEO (answer engine optimization)
The practice of increasing how often, and how accurately, a company is named or cited in answer-engine responses to its buyers' questions.
- GEO
Generative engine optimization; used by some as a synonym. We use AEO.
- Named
The company is mentioned by name in the answer text.
- Cited
A page on the company's domain is listed as a source.
- Share of voice
Named runs ÷ total runs for a prompt cluster and engine, expressed as a percentage with a 95% interval.
- Position credit
1.00 first named, 0.85 second, 0.60 third, 0.35 fourth or later, 0 not named — reported alongside, never instead of, share of voice.
- Variance
The spread of "named" across the five repeats of the same prompt; published per prompt.
- Entity facts
Checkable statements about a company: certificate, approvals, ratings with dates, fleet or capabilities, locations, contacts.
- Placement
An appearance of a company on a third-party page (list, directory, article) that an engine retrieves.
2. Prompt panel rules
- Prompts are written in the buyer's words, not ours; each names a segment, and at least one of: route, airport, aircraft type, service, rating or standard, region.
- No prompt names a company (brand-name prompts are measured in client work, not in the Index).
- Balance: seven segments and four stages, tagged by region and mission — charter operators 40, brokers 15, FBOs 20, MROs 15, management and sales 15, fuel 5, vendors and software 10.
- Frozen per quarter; changes are proposed monthly, applied only at quarter boundaries, and logged publicly with reasons.
- The full panel is public.
- Clients and subscribers may not propose prompts. Prompts come from the Buyer Question Program: consented interviews with operators, brokers, FBO and MRO managers, flight-department buyers and vendors recruited through public channels; the public opt-in form on this site; and cited secondary sources. Nothing is sourced from inside aisba.ai.
- The vendors and software segment excludes the charter sourcing, booking and payments category in Edition 1; the exclusion is lifted at a quarterly panel review only under the conditions in the conflict policy.
- Edition 1 is US-only; UK/EU and Middle East cuts arrive with Edition 3.
3. Engines, measurement surfaces and limitations
API-measured: the models' retrieval-augmented APIs, not the consumer products
Engines and measurement surfaces · ranked API engines, the calibration sample, and the client-work check
| Engine | Measurement surface | What it is not | Label used in tables | Repeats |
|---|---|---|---|---|
| Ranked · official, paid APIs | ||||
| OpenAI (GPT family)openai api + web search, ranked api engine | OpenAI Responses API with the web_search tool, default settings; user_location US unless a region cut; fresh call per run | Not the ChatGPT app: the model's retrieval-augmented API, not the consumer product. | OpenAI API + web search | 5 |
| Google Geminigemini api + search grounding, ranked api engine | Gemini API with Google Search grounding; fresh call per run(verbatim text published only if Google's grounding terms permit storage and display; otherwise derived data only) | Not the Gemini app or Google AI Mode: the model's retrieval-augmented API, not the consumer product. | Gemini API + Search grounding | 5 |
| Perplexityperplexity sonar api, ranked api engine | Sonar API; fresh call per run | Not the Perplexity app: the model's retrieval-augmented API, not the consumer product. | Perplexity Sonar API | 5 |
| Anthropic Claudeclaude api + web search, ranked api engine | Claude API with the web search tool; fresh call per run | Not the Claude app: the model's retrieval-augmented API, not the consumer product. | Claude API + web search | 5 |
| Not ranked · hand-run calibration sample | ||||
| Google AI Overviews / AI Modeai overviews (manual sample), hand-run sample, not ranked | no official API; quarterly hand-run calibration sample, 20 prompts × 2 runs entered by an analyst in a clean US browser profile, logged with screenshots | Not measured: sampled. Reported as an agreement rate with the API engines, never as a ranking. | AI Overviews (manual sample, n=40) | 2 per quarter (sample) |
| Microsoft Copilotcopilot (manual sample), hand-run sample, not ranked | no official API; same calibration sample. Candidate for Edition 2: "Azure OpenAI + Grounding with Bing Search" as an additional API-measured engine, labeled as a Bing-grounded proxy, not as Copilot | Not measured: sampled. Enterprise Copilot instances cannot be observed from outside; a site owner sees only Bing Webmaster Tools' sampled, aggregate count of citations of its own pages. | Copilot (manual sample, n=40) | 2 per quarter (sample) |
| ChatGPT consumer appchatgpt app (manual sample), hand-run sample, not ranked | included in the calibration sample so readers can see how far API results diverge from the app | Not the API-measured OpenAI surface: the app, sampled by hand, never ranked. | ChatGPT app (manual sample, n=40) | 2 per quarter (sample) |
| Client work only · never in the Index | ||||
| Client work only: Google AI Overviewsaio (licensed check), licensed check, client work only | a licensed Google AI Overviews check on a client's own prompts, supplied by a licensed tool that will be named here, with its terms, before the first client Diagnostic; shown in its own column, never in the Index | Not part of the Index. Never pooled with API runs. | AI Overviews (licensed check) | per engagement |
Measured through official APIs OpenAI API + web search · Gemini API + Search grounding · Perplexity Sonar API · Claude API + web search — the models' retrieval-augmented APIs, not the consumer products. How each surface is measured
Counsel's review of the four API terms and the consumer-app terms is pending; until it is complete, every engine is treated as derived data only.
There is no official API for Google AI Overviews, AI Mode or Microsoft Copilot; every tracking tool obtains those answers by scraping Google's or Microsoft's consumer surfaces, and Microsoft's "Grounding with Bing Search" returns an Azure model's answer, not Copilot's. We will not scrape consumer apps, and we will not launder scraping through a vendor for a published benchmark. The public Index therefore ranks only engines we can measure through official, paid APIs, and it says on every table that it measures the models' retrieval-augmented APIs, not the consumer products. API answers differ from the apps — different system prompts, model routing, no personalization, no logged-in state. That is a limitation we publish, and it is why the Index is reproducible: anyone with API keys can re-run the panel. Manual sampling by a person using a product normally is not automated scraping; whether the consumer terms permit publishing screenshots or quotations is part of the counsel review, and if they do not, the calibration sample is reported as agreement statistics only. Every number on this site names the surface it was measured on. Enterprise Copilot and ChatGPT Enterprise instances, where much B2B research happens, cannot be observed from outside. 1 Since February 2026 a site owner can see only Microsoft's aggregate, sampled count of citations of its own pages in Bing Webmaster Tools, which shows no answers, prompts or tenants. 23
Limitations
- Prompt wording is ours; a differently worded panel would produce different lists.
- The Index measures retrieval-augmented APIs, not the consumer apps: different system prompts, model routing, no personalization, no logged-in state. The calibration sample shows how far they diverge; it does not close the gap.
- Google AI Overviews, AI Mode, Copilot and the ChatGPT app are not ranked, because they cannot be measured reproducibly without scraping.
- Enterprise Copilot and ChatGPT Enterprise instances, where much B2B research happens, cannot be observed from outside. Since February 2026 a site owner can see only Microsoft's aggregate, sampled count of citations of its own pages in Bing Webmaster Tools, which shows no answers, prompts or tenants. 23
- Edition 1 is US-only; region is fixed per run.
- Five repeats distinguish patterns from noise; they do not remove it.
- Being named is not being chosen.
The instant check at /check is not the Index. It runs each of its three prompts once through the official APIs of Perplexity and OpenAI, so it says nothing about run-to-run variance, and it reads the API answers, not the consumer apps. The "companies named" lists are extracted by software and not confirmed by a person: a capitalized phrase counts as a company only when it is not an engine, a place, a standards body or an aircraft type, and a phrase that is an aircraft type plus Jet, Jets, Air or Aircraft is dropped too. That rule has known casualties: real companies whose names have that shape, such as "Citation Air", "Falcon Air" or "Global Jet", can be missing from the list. The free Snapshot has a person confirm every named / not-named decision.
4. Runs, cadence and logging
- Each prompt is run five times per engine per month through the API, spread across the month, each as a fresh call with no conversation history; the calibration sample is entered by hand in a clean browser profile in the last month of each quarter. Each run is logged with: prompt id, engine, model version, API parameters, region, timestamp (UTC), answer text, listed sources (URLs), the entities we extracted, and the terms-of-service version in force.
- Named entities are extracted with software assistance; every named / not-named decision is confirmed by a named analyst, and any prompt where repeats or reviewers disagree goes to a second reviewer. No automated attribution is published unreviewed.
- Raw answers are retained for 24 months and then deleted; derived data is kept. With each edition we publish the derived run log; verbatim answer text only for engines whose terms permit publication (confirmed per engine by counsel before this page went live — until then, derived data only); names of individuals are redacted by a pattern pass and an analyst review before anything is published, and UK/EU cuts treat sole-trader brokers as personal data. The license is stated on each edition; our intention is CC BY 4.0 for the derived data and aggregate tables.
5. Scoring
Share of voice is named runs ÷ total runs for a prompt cluster and engine, expressed as a percentage; the interval is a Wilson score interval at 95%.
Position credit
| Position in the answer | Credit |
|---|---|
| First named | 1.00 |
| Second | 0.85 |
| Third | 0.60 |
| Fourth or later | 0.35 |
| Not named | 0 |
We publish share of voice first and position credit second, and never a single "rank" number without its interval.
| Company | Named in | Share of voice | 95% interval |
|---|---|---|---|
| Company A | 5 of 5 | 100% | 57%–100% |
| Company B | 3 of 5 | 60% | 23%–88% |
| Company C | 1 of 5 | 20% | 4%–62% |
Illustrative — not measured data
Also defined here: citation share and accuracy flags, as on the Index page.
6. Attribution rules
- Hard match: a source URL on the company's registered domain(s).
- Soft match: the company name plus a distinguishing fact (base, type) in the answer text; ambiguous names (for example a company sharing a name with an airport or a person) require two distinguishing facts.
- Groups and subsidiaries are attributed to the operating brand named in the answer, not the parent.
- A company named only as a negative example ("avoid X") is recorded as named-negative and excluded from share of voice.
“Which Part 145 shops do heavy checks on a Challenger 604 in the Southeast US?”
Two shops came up: Shop A, whose own capability page lists the Challenger 604 heavy check with typical lead times, and Shop B at KOPF, an authorized service center on the OEM's list. No other shop is mentioned on a current page.
Not in the answer: Shop C not named
- Shop A — named and cited — hard match: a source URL on its own domain
- Shop B — named, not cited — soft match: the name plus a distinguishing fact (its base); the source is the OEM's page, not its own
- Shop C — not named — a Part 145 shop in the region with the rating that this run did not mention; recorded as not named, never inferred
- Source 1: shop-a.example/capabilities
- Source 2: OEM authorized service centers page
- Source 3: pilot forum thread (illustrative)
7. What we will not claim
The promise block from Services, reused verbatim, plus three sentences that apply to everything we publish.
What we promise, and what we do not
We promise
- Logged, repeatable measurement with the run data (where provider terms permit).
- Named people doing the work.
- Published prices.
- A three-month initial term and then month to month.
- A written monthly report you could hand to a skeptic.
We do not promise
Any rank, any traffic figure, any number of inquiries, or that any engine will name you. AI answers vary by engine, account, region and day, and we measure that variance rather than hide it. 4
- We do not publish before/after results from one run.
- We do not publish client results without the same panel and method used before and after.
- We do not use "AI visibility score" as a proprietary number; we publish the components.
8. Crawler policy (robots.txt) and crawler status
We allow every answer-engine crawler, including training crawlers, because a practice that sells AI visibility should be in the training data. Here is our robots.txt and what each agent does.
Answer-engine crawlers · what each one does and what blocking it costs
| Agent | Operator | Purpose | Effect of blocking | Our setting |
|---|---|---|---|---|
| OAI-SearchBot | OpenAI | ChatGPT search index | "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers" | Allow |
| ChatGPT-User | OpenAI | User-initiated fetch | "robots.txt rules may not apply" | Allow |
| GPTBot | OpenAI | Training | Excludes from training; "does not affect ChatGPT search results" | AllowA practice that sells AI visibility should be in the training data. |
| Claude-SearchBot | Anthropic | Claude search relevance | "May reduce visibility in Claude's search results" | Allow |
| Claude-User | Anthropic | User-initiated fetch | Prevents retrieval in Claude conversations | Allow |
| ClaudeBot | Anthropic | Training | Excluded from future training | AllowCrawl-delay is supported if ever needed. |
| PerplexityBot | Perplexity | Perplexity search | "To ensure your site appears in search results, we recommend allowing PerplexityBot" | Allow |
| Perplexity-User | Perplexity | User fetch | "generally ignores robots.txt" | Allow |
| Googlebot | Search, AI Overviews, AI Mode | Removes from everything | Allow | |
| Google-Extended | Gemini training and grounding control token; "does not impact a site's inclusion in Google Search nor is it used as a ranking signal"; no separate user agent | Blocks Gemini training and grounding use | Allow | |
| Applebot | Apple | Siri, Spotlight and Safari search | Removes from Apple search surfaces | Allow |
| Applebot-Extended | Apple | Training control only; "Webpages that disallow Applebot-Extended can still be included in search results" | Excludes from Apple foundation-model training | Allow |
| CCBot | Common Crawl | Open crawl used for AI training | Excluded from Common Crawl datasets | Allow |
| Meta-ExternalAgent | Meta | Meta AI; "crawls the web for use cases such as training foundation AI models or improving products by indexing content directly" | Excludes from Meta AI training and direct indexing | Allow |
| Meta-ExternalFetcher | Meta | User-initiated fetch; "fetches individual links at a user's request" | "this crawler may bypass robots.txt rules" | Allow |
| Bingbot | Microsoft | Bing and Copilot | Removes from Copilot answers | Allow |
For clients we usually recommend allowing answering and search bots (OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot, Bingbot) and deciding on training bots (GPTBot, ClaudeBot, CCBot, Google-Extended, Applebot-Extended) as a policy question; blocking training bots does not remove you from ChatGPT search. 5
Our robots.txt
User-Agent: * Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: OAI-SearchBot Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: ChatGPT-User Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: GPTBot Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: Claude-SearchBot Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: Claude-User Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: ClaudeBot Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: PerplexityBot Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: Perplexity-User Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: Googlebot Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: Google-Extended Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: Applebot Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: Applebot-Extended Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: CCBot Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: Meta-ExternalAgent Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: Meta-ExternalFetcher Allow: / Disallow: /dev/ Disallow: /api/ User-Agent: Bingbot Allow: / Disallow: /dev/ Disallow: /api/ Sitemap: https://aisaysyou.com/sitemap.xml
Which answer-engine crawlers visited aisaysyou.com · last 30 days
No log data yet. We started collecting server logs on 25 September 2026; verified crawler visits will be listed here monthly.
9. On llms.txt and structured data
We publish an llms.txt because it is free and clients ask; we do not sell it. In a 137,000-domain study, 97% of llms.txt files were never requested, and AI retrieval bots were about 1% of requests. 6 Structured data is hygiene for entity clarity (Organization, Service, Person, Dataset); FAQ and HowTo markup showed no relationship to AI Overview citation in Seer's 2026 analysis. 7 Google states there are no additional requirements to appear in AI Overviews or AI Mode.
10. The Citation Ledger (how client work runs)
Client engagements run on one ledger: prompt panel → logged runs → gaps → fixes → placements → re-runs, visible to the client in a shared dashboard with run logs downloadable (derived data always; verbatim text where the engine's terms allow). Client work may add a licensed Google AI Overviews check in its own column, named above; the Index never does. Each step uses the definitions and rules above; a client's before/after is only ever the same panel, the same surfaces and the same number of repeats.
11. Sources we rely on
Every third-party statistic on this page is listed below with its publisher, date and URL. The list is computed from the citation marks on the page; nothing is hand-maintained.
12. Change log
- v1.0.226 Sep 2026Limitations gain a paragraph on the instant check at /check: one run per prompt through two official APIs, names extracted by software and not confirmed, and the known casualties of the aircraft-type rule (company names shaped like "Citation Air", "Falcon Air" or "Global Jet" can be dropped from the companies-named list). The Index measurement does not change.
- v1.0.125 Sep 2026Copilot limitation narrowed: enterprise Copilot instances cannot be observed from outside (not "at all"). Since February 2026 Bing Webmaster Tools' AI Performance report shows a site owner a sampled, aggregate count of citations of its own pages, with no answers, prompts or tenants. Nothing in the measurement or the scoring changes.
- v1.024 Sep 2026First published.