How LoudPixel measures AI citations

This page explains the methodology behind every audit number LoudPixel shows you — where the queries come from, how the six AI engines are interrogated, what counts as a citation, how duplicates are collapsed, and how it all rolls up to a single percentage. It is written for the person who has to defend the score to a CFO or a skeptical head of growth. After reading you will know what a Citation Share Score of 18 means, what it does not mean, and why it changed week over week.

If you have not run an audit yet, start with Getting started — your first AI search audit. This page assumes you have a report on screen.

The five-stage pipeline

Every LoudPixel audit runs the same five stages, in order. The output of each stage is auditable in the UI — you can click any number on the report and trace it back to the raw API response that produced it.

1. Query generation

LoudPixel does not ask you to write your own queries — we derive them automatically from your domain and your declared competitors. The result is 20 buyer-intent queries on the free tier: things a real prospect would type into ChatGPT when they are looking to buy, not when they already know your brand name.

Branded queries ("is acme.com any good") are excluded by default. You already win those — the prize is the category-level questions you are losing to competitors or to third-party listicles.

2. Engine querying

The 20 queries are fired in parallel against six engines — ChatGPT, Perplexity, Claude, Gemini, Grok, and Mistral — through their official APIs. We do not scrape the consumer web UIs: terms of service prohibit it, and logged-in answers are personalized in ways an audit cannot reproduce. API responses are what a developer integrating the engine gets — the closest approximation to "what the model thinks today."

If a single engine errors or times out (usually Grok during X traffic spikes), the audit completes with the remaining engines and the missing one is flagged. The Citation Share Score is computed only against engines that actually answered.

3. Citation parsing

A "citation" is detected one of two ways. Either (a) your domain appears verbatim in the answer text — for example, "…tools like acme.com and competitor.com…" — or (b) your domain appears in the structured sources array some engines return alongside the answer (Perplexity and Gemini expose this; ChatGPT exposes it for some models). Either path counts as one citation for that engine on that query.

One edge case worth knowing: subdomains roll up. A citation of blog.acme.com or app.acme.com counts as a citation of acme.com — buyers recognize the brand, not the subdomain.

4. Deduplication

If three engines cite your domain on the same query, that counts as three engine-citations but as one citation in your unique-query coverage. The same domain cited twice within a single answer counts as one. The score measures coverage of the answer space, not raw volume. Inflating the count for redundancy would reward brands that get repeated within a single answer over brands cited on a wider spread of questions — the opposite of what a buyer experiences.

5. Aggregation

The per-engine, per-query matrix rolls up to a single number — the Citation Share Score — the percentage of audited (engine × query) cells in which your domain was cited at least once. The exact arithmetic and missing-engine handling is in Reference — Citation Share Score formula. One headline number wins because founders and CFOs will not stare at a ten-metric dashboard every Monday.

Why this approach

Why buyer-intent queries, not branded. A branded-query audit tells you the engines know you exist. They almost certainly do. The revenue-relevant question is whether they recommend you when a buyer asks the category question — "best CRM for small B2B SaaS", not "is acme.com any good".

Why six engines, not just ChatGPT. Citation behavior diverges sharply across engines. Perplexity over-indexes on recent blog content. Claude pulls from primary sources and documentation. Grok pulls heavily from X threads. Gemini leans on Google's index. Optimizing for one engine and assuming the others follow is a category error — they do not follow.

Why a percentage, not a count. A raw count (23 citations) does not normalize across audit sizes. Five of ten and twenty-five of fifty are the same answer, but only one is comparable across audits run with different query budgets. Expressing the result as % of cells cited keeps the 20-query and 200-query audits on the same axis.

Known limitations

• A single audit is a snapshot. Engines re-rank weekly, sometimes daily. Use the weekly digest for trend; do not treat one Monday's number as a forecast.
• Free-tier noise is real. Twenty queries shows direction, not single-percentage-point movements. The paid tier expands to 200 queries, compressing the noise band.
• Some cited content is paywalled. Engines occasionally cite sources outside the public-web audit. We log and flag these separately rather than hiding them.
• Citation does not mean recommendation. A citation may be neutral, comparative, or mildly negative. Sentiment is tracked on the paid tier; it does not enter the Citation Share Score.

What this score lets you DO

Once you can trust the number, you can act on it. The gap between your score and your top competitor's, multiplied by the average AI-referred session value for your category, gives the directional dollar figure on every report — the revenue moving through answers you are not in. Week-over-week deltas tell you whether your last content shipment moved the needle. To turn diagnosis into fix, jump to How to get cited on a missed query — the one playbook that pays back the audit fastest.