artigo

AI citation measurement protocol

How to collect, log and calculate citation share in a repeatable way: query set, platforms, repeated runs, logging, metrics, confidence intervals, monthly cadence and the pitfalls that break a series.

short answerThe Citation Share protocol, by Camillo Dantas, has nine steps: define the scope, freeze 50 to 100 queries, fix the collection environment, run each query at least three times per platform, log every answer, classify with written rules, calculate rates with a 95 percent confidence interval, repeat at a fixed interval and publish with the limits. Each step has a success criterion that another person can check.
Neste texto

Who is this protocol for?

For anyone who needs to track how often AI answers cite a brand in a way that stays comparable over time: a marketing team, an agency, a consultant or a researcher. The level is intermediate. You need to know the market and the competitors, work with a spreadsheet and have access to the platforms you want to measure. A round with 60 queries, three runs each and three platforms takes a few hours of manual work, not counting the preparation of the set.

The definition of the metric is in What is AI citation share?. The Portuguese version of this protocol, with the full tables and more cases, is Protocolo de medição.

Step 1: how do you set the scope?

Answer five questions in writing before you open any AI: which brand and which domains represent it; which three to eight competitors are the reference; who is the audience and in which language; in which country or city do customers ask; and which platforms matter to that audience.

Success criterion: one paragraph, signed by the person in charge, with the list of brand domains, including subdomains and campaign domains. Without that list there is no clear rule for classifying a citation.

Step 2: how do you build the query set?

The set holds 50 to 100 questions that customers really ask. Spread them across the decision journey: discovery, consideration, decision, comparison with competitors, local and problem-solving. Include one group with the brand name (to measure reputation and accuracy) and one group without it (to measure whether the brand enters on merit). Source them from site search, Search Console queries, support tickets, sales calls and community threads. Number them from Q001 and record the date.

Success criterion: a numbered set of at least 50 queries, with 30 percent or more being decision and comparison queries, and frozen text. Any later change becomes a new query with a new ID, because changing a word can make the question easier and the share rise for the wrong reason.

Step 3: how do you prepare the collection environment?

The environment must be as similar as possible between runs. Write it down and keep it:

  • a private window or a clean browser profile, without extensions that alter the page;
  • the same interface language and the same region (for local queries, write the city in the query);
  • a new conversation for every query, with no history;
  • logged in or logged out, but always the same, because logged-in accounts may carry memory and personalization;
  • web search on, when the platform lets you choose.

The last point matters because platforms decide when to search. OpenAI's API documentation says the model can choose to search or not, based on the prompt. Anthropic's web search tool documentation says Claude determines when to search. An answer without sources may simply mean that the model did not search, so record whether search was active.

Success criterion: a configuration sheet per platform with browser, language, region, login, search mode and the model name or version shown on screen.

Step 4: how do you run the baseline on each platform?

Run each query at least three times in new conversations, with the text pasted unchanged. Do not rephrase a poor answer: a poor answer is part of the sample. Spacing the three runs across different hours or days helps to reveal variation.

PlatformWhat its documentation says (read on October 22, 2026)What to log
ChatGPT and the OpenAI APIthe model can choose whether to search; by default it includes inline citations for URLs from the search resultswhether it searched, the cited sources, the model version
Claudein the API search tool, Claude decides when to search and citations are always enabledwhether search was active, the cited sources
Perplexityeach answer includes numbered citations linking to the original sourcesthe numbered source list and the mode used
Copilot Search in Bingcites its sources prominently, with inline links and a list of every link usedinline sources and the final list
Google AI Overviews and AI Modemay use several related searches; AI feature traffic is counted in the Search Console Performance report under the Web search typewhether the feature appeared; if it did not, mark "not applicable"

You do not need to measure every platform. Pick the ones your audience uses. Three platforms measured well are better than seven measured thinly.

Success criterion: for every query and platform, at least three saved answers with date and time, or a written reason for the gap.

Step 5: how do you log each answer?

Log one row per answer at the moment of collection. The minimum fields are date, platform, mode or version, query ID, run number, status (valid or invalid), brand mentioned, brand domain cited, brand recommended, information correct, competitors cited and source domains. Keep the full answer text or a screenshot with the date visible, for audit. The benchmark sheet has this structure and exports CSV.

Success criterion: another person can check any row against the original text without asking the collector anything.

Step 6: how do you classify without bias?

Classification is the most subjective step. Written rules reduce disagreement between reviewers.

FieldMark "yes" whenEdge cases
Valid answerthe platform gave a complete answer to a valid questionerrors, refusals, cut-off or empty answers are invalid and leave the calculation
Brand mentionedthe brand name appears in the answer texta name inside a list of "other options" is a mention, not a recommendation
Domain citeda URL on the brand's domain or subdomain appears among the sourcesthird-party text about the brand is not a citation of the domain; note it as an observation
Recommendedthe answer points to the brand as a suitable option (decision queries only)listing the brand without an assessment is a mention
Information correctwhat the answer says matches the brand's official pagea partly right answer is "no", with the error described

When two reviewers classify the same 20 answers, disagreement should stay under 10 percent. If it does not, rewrite the rule of the field that diverged before going on.

Step 7: how do you calculate and read the rates?

Calculate per platform. Always read three things together: the count (29 of 118), the rate (24.6 percent) and the 95 percent confidence interval (17.7 to 33.1 percent). Those numbers come from a fictitious example. The method uses the Wilson interval, and the calculator applies it.

Sample size drives the width. For an observed rate near 25 percent, the Wilson 95 percent interval is about:

Valid answersInterval
3013 to 43 percent
6016 to 37 percent
12018 to 33 percent
24020 to 31 percent
48021 to 29 percent

This treats each answer as independent, which is the best case. Because runs of the same query resemble each other, also calculate the rate with the query as the unit: a query counts as cited when the domain appears in at least half of its valid runs. Report that interval. Vendor research supports the caution: a study by SparkToro and Gumshoe.ai, published on January 28, 2026 (a vendor study, not peer reviewed), found that for ChatGPT and Google's AI answers the chance of two responses to the same prompt returning the same list of brands was under 1 in 100, and its author considered a visibility percentage across many prompts and runs more reasonable than a position.

Success criterion: a report per platform with count, rate and interval, on both units.

Step 8: when and how do you repeat it?

Repeat at a fixed interval, with the same set, the same platforms and the same configuration. Monthly or quarterly is enough in most cases. Note the events between measurements: content published, a site change, a campaign, a competitor launch and a platform model change.

To say that something changed, compare the two measurements through the interval of the difference. If the interval includes zero, the correct conclusion is that real change cannot be separated from noise. In the fictitious example, a rise from 11.3 percent (13 of 115) to 24.6 percent (29 of 118) is a difference of 13.3 points with an interval from 3.4 to 22.9: it excludes zero, and it is also wide.

Success criterion: a time series per platform, with events noted and no comparison between different configurations.

Step 9: how do you publish honestly?

Publish a percentage only with a minimum sample, the protocol described, the collection date, the platforms and versions, the query list (or at least its structure) and the limitations. Label it an "observation" and not a "position". If the study was run by the vendor of the service or by the owner of the measured brand, say so. Avoid averages across platforms and avoid comparing periods with different sets.

Success criterion: an outside reviewer can repeat the collection from the published description.

Which pitfalls break a series?

PitfallHow it shows upControl
Query selectionthe set favors what the brand already coversuse real customer questions and include hard ones
Variation between runstwo runs give different answersrepeat each query and report an interval
Personalizationa logged-in account gets answers shaped by historynew conversation, clean profile, log the login state
Location and languageanswers change by regionfix region and language, write the city in the query
Model changeresults move although the brand did notlog version and date, compare only like with like
Reviewer biaspeople who know the brand see more mentionswritten rules and a second check
Small samplewide intervals and fragile conclusionsadd queries, not only repeats
Manual versus API collectionAPI output can differ from what users seedeclare the method and do not mix them in one series

The last row echoes an open question in the SparkToro study: whether API calls match the variety of real user sessions is unresolved.

What is the next step?

Build the set with the query generator, log the first round in the sheet and calculate it in the calculator. For the full routine, including diagnosis and an action plan, the course Citation Share: measure and prove costs R$ 397. Measuring does not promise that any brand will be cited.

How many queries do I need to start?

The protocol recommends 50 to 100. With fewer than 30 valid answers per platform, the interval is wider than 30 percentage points and the result is exploratory. You can begin smaller to learn the routine, as long as you label the result a pilot.

Can I automate the collection?

Yes, if you declare the method. API output can differ from what a user sees in the interface, and the question is still open in the market. Do not mix manual and automated collection in one series, and log the model version and parameters.

What do I do when an answer shows no sources?

Log it as valid, mark the domain as not cited and note that there were no sources. If search was off or not triggered, log that too. The rate then mixes answers with and without sources, and the note helps with the reading.

How often should the classification rules change?

Almost never. Changing a rule in the middle of a series breaks the comparison. If it is unavoidable, version the rules and reclassify the old measurements, or treat the change as the start of a new series.

Sources

  1. Web search, OpenAI API docs (2026-10-22)
  2. Web search tool, Claude Platform docs (Anthropic) (2026-10-22)
  3. How does Perplexity work?, Perplexity Help Center (2026-10-22)
  4. Introducing Copilot Search in Bing, Microsoft Bing Blog (2025-04-04)
  5. AI features and your website, Google Search Central (2025-12-10)
  6. AIs are highly inconsistent when recommending brands or products (vendor study, not peer reviewed), SparkToro and Gumshoe.ai (2026-01-28)

Meça antes de otimizar

Defina o conjunto de consultas, registre a linha de base e acompanhe a evolução com o mesmo protocolo.