# How to measure what AI says about you, free · The method · EarnCitations

> A step-by-step method for measuring how AI assistants answer your buyers: fixed questions on a schedule, claims checked against sources, fixes confirmed. Free tools.

Source: https://earncitations.com/method.html

> Every page on this site: https://earncitations.com/llms.txt (the whole site as one file: https://earncitations.com/llms-full.txt)

Method updated 2026-11-22

## How to measure what AI says about you, step by step.

You don't need a paid tracker to learn whether ChatGPT, Perplexity or Google's AI answers name you, recommend you and get your facts right. You need fixed questions, a schedule, a careful way to count, and a record. This is the method we follow, with the free tool for each step.

Much of it follows the measurement method Jam (spreadjam.com) has published, and we credit it where we use it. Verify delivery before reading answers. Count per answer, not per mention. Keep the kinds of presence apart. Classify an error before correcting it. Where we differ, we say so. The commands shown run from a copy of the repository ([quickstart](https://earncitations.com/developers.html#quickstart)); every step also has a free page here.

### 1. Write the questions down, then freeze them

List the questions a buyer would ask before choosing you: what it costs, whether it does X, how it compares, who it's for. Split them in two. Discovery questions don't name you ("best CRM for small agencies?"), and they show whether buyers find you. Diagnostic questions do name you ("does Acme have SSO?"), and they show whether what's said about you is right. Report the two apart: a high rate on diagnostic questions says nothing about discovery.

Freeze the wording. A reworded question is a new question, so compare its answers separately. The [question map](https://earncitations.com/map.html) drafts questions from your own pages and exports them as a versioned question set (earncitations map you.com --set questions.json).

Keep one wording per buying decision. "Best CRM for startups" and "top CRM tool for startups?" are one decision asked twice, and tracking both makes that decision count double. The answer log points out wordings that share almost all their words when you load a set. A handful of questions that stand for real decisions beats hundreds of variations.

### 2. Make sure crawlers can read your answers

An assistant can only use a page its crawler could fetch and read. Run the [audit](https://earncitations.com/#audit) for robots.txt rules per AI agent, pages that need JavaScript to show text, missing structured data and llms.txt. The [crawler directory](https://earncitations.com/crawlers.html) lists which agent does what, from each operator's own documentation.

### 3. Ask on a schedule, and keep everything

Ask each question on each engine you care about, in a fresh chat, signed out or with memory off where you can. Note the engine and mode ("ChatGPT search", "Perplexity Pro"). Keep the exact answer, its links, the question as asked, and when. Ask the same set every round: earncitations track due answers.jsonl --set questions.json --engine ChatGPT,Perplexity --days 7 lists what's still to ask.

Answers vary from one try to the next. Several answers to one question on one day are repeats, not more buyers, so the answer log flags them.

### 4. Count per answer, four ways, and name your competitors

Paste each answer into the [Answer check](https://earncitations.com/claims.html) with your brand and domain (or select it and use the Answer check's bookmarklet), then save it to the [answer log](https://earncitations.com/track.html). For every answer we record four things that overlap and so are never added or subtracted:

- Named: the answer mentions you. Five mentions in one answer count once.

- Recommended: a judgment favours you: "best for…", "our top pick", "a strong alternative", "a solid choice", "worth considering". "X is better than you", "we don't recommend you" and a bare "strong security" don't count.

- Your page cited: a link the answer gives is on your domain. Cited but not named is source visibility, not a recommendation.

- Described by others: a claim about you that is stated only on another site's page.

Name up to 10 competitors and each gets the same reading: named, recommended, their own page cited, and whether they come before you. Every rate shows its denominator.

Count cited links per page as well as per site. The answer log lists which of your pages answers cite and which other pages they lean on, each counted once per answer. It removes campaign parameters and #sections and keeps other query strings, since ?version=2 can be a different document. A domain total can't tell your comparison page from an outdated tutorial.

### 5. Check every claim against the pages it cites

The Answer check splits an answer into claims and reads each cited page the way a crawler does. Each claim is supported (a page says it, with every price, duration and percentage matching), contradicted (a passage about the same thing gives a different figure), unaddressed or unknown (the page couldn't be read). "Supported" means a cited page says it, not that it's true. "The page doesn't say it" isn't "it's false".

Then classify what's wrong before fixing it: incorrect, outdated, the wrong company, unsupported, a subjective view, or an accurate limitation. Each needs a different response. The ledger download (CSV, one row per claim) has columns for that.

### 6. Decide what to do

Each problem gets one of four responses: explain something you already do more clearly, correct a factual error at its source, accept a real difference (a product gap isn't a content gap), or investigate further. The [question doctor](https://earncitations.com/question.html) finds the page that should answer a question and, when your site already states the answer somewhere, drafts the fix from your own words. When another site's page contradicts yours about you, the Answer check drafts a correction request quoting both pages. You check it and send it yourself, and their opinion stays theirs. The answer log lists the other sites your answers cite most: those are the sources assistants already trust for your questions, and where a mention or a correction counts.

Turn the audit's failing checks into work someone owns: the audit result downloads them as a backlog (CSV) or copies them as an issue task list, most important first, each with the kind of role that usually does it and how you'll know it's done.

### 7. Ship the fix, and confirm crawlers got it

A fix that never reached a crawler can't change an answer, and reading answers a week later can't tell "it didn't work" from "it never arrived". The [Delivery check](https://earncitations.com/delivery.html) confirms, step by step, that the page is live, has the new text in its HTML without JavaScript, is allowed in robots.txt, is indexable, is in the sitemap with an accurate lastmod, and is linked. earncitations indexnow submit tells Bing, Yandex and the other IndexNow engines that the page changed.

### 8. Re-ask, and read the change carefully

Give crawlers time to come back (how long varies; assistants that fetch a page live when asked can see it sooner), then ask the same frozen questions again. On the answer log, "Before and after a change" takes the day the fix went live and gives each question on each engine one of four outcomes: improved, unchanged, worse or inconclusive (fewer than three answers on two or more days on either side). On the command line it's earncitations track compare answers.jsonl --change 2026-11-20. A change that follows an edit is a sequence, not proof the edit caused it: other pages, model updates and ordinary variation move answers too. Look for the same change across questions and engines before believing it, and check its size against the chance swing the [answer log](https://earncitations.com/track.html#due) gives for your routine: with about 170 answers a month, a rate can move some 11 points between two months by chance alone.

Then look past mentions to what the visits do. The [log analyzer](https://earncitations.com/logs.html) counts the people who clicked through from an AI answer and reached a sign-up, demo or contact page. With no server logs, it reads a Google Analytics, Plausible or Matomo export instead: sessions from AI assistants, their key events and revenue against everyone else's, and the pages assistants send people to. These are floors, because app visits can arrive with no referrer.

### What we don't do

- No single "visibility score". The four kinds of presence measure different things, and a blend hides which one moved.

- No sentiment score. We list the claims that put you with a caveat ("but", "lacks"), and you read them.

- No claim that a change caused a result. We show the sequence and the denominators.

- No AI calls in any of the steps above. The Answer check matches words and figures, not meaning, so check the claims that matter yourself.

Everything above runs in the browser or from the command line ([CLI, API and MCP server](https://earncitations.com/developers.html)), free.
