Skip to content

Methodology

Every finding we publish is measured, not guessed.

Each finding traces back to the answers it rests on and, where a claim is checked, to the record that settles it. This page sets out how, and where the measurement stops.

What we query
Buyer-intent questions put to the developer APIs of OpenAI, Anthropic, Google and Perplexity in your market frame. What we measure
What we record
Every answer, the entities named in it, the order they appear, and the sources cited. How answers are read
How it is checked
Against the primary record a claim rests on, with an interval on every reading rather than a bare number. How we verify
What it cannot establish
It measures answers, not prescribing, sales or patient outcomes. Where the measurement stops
What evidence assurance means
Evidence assurance is a defined check. For a claim in an AI answer, it identifies the record competent to settle it, in that market on that date, tests the claim against that record, and records the result and its limits. It covers the claims in scope, not every claim in an answer and not the answer's reasoning. It is not a certification, an audit opinion or a guarantee that any answer is correct, and it does not replace your own review, including medical, legal and regulatory review. What it covers and where it stops
Last updated
14 November 2026
Sections
12
Stated boundaries
4

Operating stance

Foundational accuracy and neutrality are the product. AlphaCitation measures what AI systems tell people, where those answers come from, and whether the documented record supports them. The method is the same for every client in every sector, and no client can change it: the questions are set before measurement begins, the engines and the decision rule are fixed, and every finding traces back to the verbatim answers behind it and, where a claim is checked, to the primary source it was checked against.

We report what we measure, including results a client would rather not see. We help authoritative information become findable and verifiable, and we never claim to control what any AI system says. We do not seed content, fabricate grassroots activity, produce synthetic media, or profile or target individuals. We work for companies, institutions and campaigns on these terms and no others.

What we measure

Each scan puts buyer-intent questions, the kind a real prospect would ask an AI assistant, to the developer APIs of OpenAI, Anthropic, Google and Perplexity, reported as ChatGPT, Claude, Gemini and Perplexity. We record whether your brand is named, where it ranks against competitors, and how it is framed.

Precision, stated

Each reading is measured until it is precise enough to report. Where that precision cannot be reached, the reading is published with its wider interval, never presented as though it had been.

Confidence intervals, not point estimates

Every measured reading carries a 95% confidence interval rather than a bare number. Where the uncertainty cannot be pinned down, the interval says so, rather than rounding to look more precise than the data supports.

Deterministic reading

Given the same answers, the reading always produces the same result. There is no hidden randomness in how an answer is read once it has been collected.

What this does not do

Every reading is a point in time. AI answers shift as the underlying models and their sources change, so a later scan can move for reasons unrelated to anything you did, which is why movement across scans matters more than any single reading. Our own reading of each answer is not assumed to be perfect.

Data sources

Every scan queries the developer APIs of OpenAI, Anthropic, Google and Perplexity directly, never scraped. Questions about your brand and its competitors are asked fresh on every scan; questions shared by every brand in a category may be reused for up to 12 hours. Some plans add sentiment from public platforms, reported as a separate, labelled signal and never blended into the AI-engine readings.

Findings are checked against official records rather than our own database: for medicines, the FDA's labels and the EU product information the EMA publishes for centrally authorised medicines, each deciding a claim only in the market its regulator governs; for other sectors, the authoritative public register, among them NHTSA safety ratings, EU Ecolabel and EU eAmbrosia. For supplements and packaged foods, the NIH Dietary Supplement Label Database holds labels transcribed by the NIH from dietary supplements sold in the US, including products no longer on the market, and USDA FoodData Central holds label data supplied by food industry data providers; a label on file in either is a record of what the label says, never a registration or a regulator's approval. A source that returns nothing is recorded as absence of confirmation, never as evidence of absence.

Controlled measurement environment

Measurement runs in a controlled, repeatable query environment against each provider's model interface. That is what makes like-for-like comparison possible: the same question design, the same reading and the same decision rule, cycle after cycle, with every question asked recorded verbatim, so a change in the reading is a change in the provider rather than a change in how we asked.

An individual consumer session can differ. Personalisation, interface-specific retrieval, search behaviour, model routing, account configuration and product-level system instructions all vary between users and between apps. Each measured response is therefore an observation within a defined test environment, not a claim that every result reproduces every possible consumer session. What the method is built to do is measure provider behaviour, and movement in it, repeatably over time.

Observed consumer surfaces

AlphaCitation captures Google AI Mode and AI Overview answers observed by people in their markets, under a pre-registered, timestamped protocol. These observations are analysed alongside API measurements, with consequential claims checked against the authoritative evidence applicable to each market. Model APIs tell us what AI systems return; Google AI Mode and AI Overviews are what people meet when they search. Each observation is preserved as it was shown, with its sources and context, and timestamped.

For the published study of 14 November 2026, the observer was in the United Arab Emirates, using US and Irish search settings. Observations are reported beside the API measurements and never pooled with them. An observation carries no confidence interval and is not continuous monitoring: it shows what a surface showed a person, at a stated time.

Asked and read in the market’s own language

An English answer translated afterwards is not the answer a clinician in that market reads. AlphaCitation asks in the market's language and reads the answer as written. Before a language is used, our reading of it is held to the same accuracy bar as English. Local languages, including Arabic, French, Spanish*, German and Asian languages. *In testing on live engine answers.

Verifying what we found

For medicines, Label Ledger checks claims about a medicine's approved uses and populations against the current label, dated, that governs the market: the FDA's in the US, the EU product information of centrally authorised medicines in the EU and EEA, and both as benchmarks elsewhere; findings that rest instead on claims the client registers itself are labelled as such. Which evidence can settle which kind of claim, for which market and as of which date, is set out in our open standard, Which Evidence Settles Which Claim (version 3.0, doi:10.5281/zenodo.23594909). The answers behind the findings are stored with a hash of their content, linked to the record before them, and once a day, if the record has changed since its last timestamp, its latest state is timestamped by an independent authority (RFC 3161). See Label Ledger.

Where the measurement stops

Where the measurement stops

A reading is only trustworthy if its edges are stated. Four boundaries describe where this measurement stops.

  1. Where a reading needs history

    One cycle shows where something stands; showing movement takes the same thing observed on more than one occasion.

    Some readings need the same thing observed on more than one occasion, or need a source's own revision history alongside what the answers say. A single cycle establishes where something stands, not how it moved. Where movement is reported, it is movement in the answers, measured between cycles this programme ran.

  2. Where a reading needs a consequence

    We observe AI answers and the evidence behind them, never prescribing, dispensing, patient outcomes or a regulator's conclusion.

    This programme observes what AI says and whether the evidence supports it. It observes no prescribing, no dispensing and no patient outcome, and it makes no regulatory determination. What an answer caused, and what a regulator would conclude about it, sit outside what any scan of answers can establish.

  3. Where a reading needs a record we do not read

    A comparison that needs a record outside AI answers is made only where that record is public.

    Some comparisons need a record that lives outside AI answers: a payer's own determination, a manufacturer's communication, a fact about your organisation. Where such a record is public, the comparison is made and reported on the surface that holds it. Where none is public, no accuracy reading is claimed rather than estimated.

  4. Where a reading is made in aggregate

    Some readings are taken across a whole question battery or at the level of a claim, rather than item by item.

    Several readings are taken across a whole question battery, or at the level of a claim, rather than item by item. A finer cut of the same data is a different judgement rather than a smaller one, so it is named here instead of implied.

Questions about this methodology? analyst@alphacitation.com.