How to track AI citations

Why one-off checks mislead

Model answers are non-deterministic. The same question asked twice can produce different sources. Personalisation and chat memory skew results towards brands you have discussed before, which is why founders checking their own brand almost always get a flattering answer.

Run the same question ten times and you will typically see your brand in some proportion of them. That proportion is the real metric, and you cannot see it from a single check.

Designing the probe set

Aim for 20 to 40 questions across four types: category questions, comparison questions, problem questions, and brand-name control questions. Write them in the buyer's language. Freeze the list — the value is in the trend, and changing questions resets it.

Run each question against every model that matters, in a clean context with no chat history, and record the full answer text plus the cited URLs.

What to record

Mention (was the brand named), citation (was our URL in the sources), sentiment or accuracy of the description, and the competitor set that appeared. Store the raw answer, because six weeks later you will want to know how the description changed, not just whether it existed.

Aggregate weekly. Daily aggregation is noise.

Turning it into action

Sort by the questions where you are consistently absent and a competitor is consistently present. That list, in order, is your content plan — each line names a page that needs to exist or needs to be rewritten.

Re-probe two weeks after each fix. If retrieval improved but citation did not, the page is being read and rejected, which is a quotability problem, not a coverage problem.

What we measured on real domains

Everything below comes from ThoughtForge runs against live domains, not from a study we read. Each run scores three things out of 100: identity (can an engine say who you are without contradicting itself), durability (does your brand survive a multi-turn conversation), and retrieval (can your pages be chunked into something quotable).

Across every domain we have scanned, one number is consistently the worst, and it is not the one people optimise for. Durability sits at or near zero almost everywhere. Brands that get named in turn one disappear by turn three, because nothing on the site gives the model a reason to keep holding onto them.

• thoughtforge.app, July: identity 16, durability 16, retrieval 49, overall 27.
• thoughtforge.app, September, after entity and schema work: identity 52, durability 0, retrieval 48, overall 33.
• A services domain (megap•••••••••••.com.au): identity 28, durability 0, retrieval 47, overall 25.
• An industry association (propertyasso•••••••.com.au): identity 4, durability 0, retrieval 50, overall 18.
• A consultancy (obse•••••••••.com): identity 40 in July, 52 in September; overall 30 to 35.

What a tracking baseline should look like

A useful baseline is three numbers taken the same way every time, not a feeling about whether ChatGPT liked you this week. Ours are identity, durability and retrieval, each out of 100, re-measured on a schedule so the deltas mean something.

The deltas are where the value sits. Identity 16 to 52 tells you the schema work landed. Retrieval 49 to 48 tells you the page rewrites have not happened yet. Durability 16 to 0 tells you multi-turn persistence got worse while you were looking elsewhere. A single snapshot tells you none of that.

Build the smallest tracker that tells you something

Ten questions, four engines, run monthly, recorded in a spreadsheet: mentioned or not, cited or not, which competitor appeared instead. That is enough to see direction, and direction is what you are buying.

Keep the wording of the questions fixed. Changing the prompts between runs makes the comparison meaningless, and it is the easiest way to accidentally manufacture good news.

What to do with a bad reading

Bad readings are more actionable than good ones if you separate the cause. Never cited anywhere means access or structure. Cited but described wrongly means identity. Named early then dropped means durability.

Change one thing, wait a fortnight, measure again. Changing four things at once produces a number you cannot attribute, which is how teams end up maintaining rituals that do nothing.

Frequently asked questions

Can I do this manually?

For a handful of questions, yes, in a fresh incognito session with memory disabled. Beyond about ten questions across four models it stops being practical to keep consistent.

How many runs per question are enough?

Three to five per model per week gives a usable rate without excessive cost. One run per question tells you very little.

What is a good citation rate?

It depends entirely on category competitiveness. What matters is your own trend line and your rate relative to the competitors appearing in the same answers.

Related guides

• AI search visibility — AI search visibility is how often answer engines mention and cite your brand. How to measure it properly, what moves it, and what to ignore.
• How to rank in ChatGPT — ChatGPT does not rank pages, it picks sources to cite. Here is what makes a page citable, how to measure it, and the fixes that change the answer.
• Why AI doesn't mention your brand — Six concrete reasons ChatGPT and Perplexity skip your brand, how to tell which one applies to you, and the fix for each.