How To Track Brand Mentions Across AI Models

What Changed For twenty years, finding a supplier meant typing a query and being handed a list. You compared a few results, formed your own opinion and chose. The businesses that appeared near the top of that list got most of the attention, which is why an entire industry grew up around getting there.

Nor has any of this removed the need for a real product and real customers who will say so. If anything it has increased it, since corroboration from independent sources now feeds directly into whether a machine will recommend you.

Tracking this is genuinely awkward, and pretending otherwise is how most reporting in this field goes wrong. There is no console. Answers vary between runs. Referral attribution is inconsistent between assistants. Anyone handing you a single confident number has hidden a great deal of variance behind it.

For roughly twenty years the arrangement was stable enough that an entire industry could be built on it. You typed a query, you got a ranked list, you formed your own opinion by comparing a few of the results, and businesses competed for position in that list.

The last of these is the most common and the hardest to see, because it produces no error anyone internally encounters. Your site works perfectly in every browser while returning a challenge page to every legitimate retrieval agent.

The Rendering Question This is the one real technical constraint. Content that only exists after JavaScript executes may be invisible to a retrieval fetch, which is not a browsing session and does not always run scripts.

What analytics cannot tell you is how often you were named without a click, which in this channel is most of the time. A recommendation that a buyer acts on three weeks later leaves no trace in any report you own. This is why the manual prompt set is not optional, and why nobody should be asked to justify this work on referral traffic alone.

The fix is not abandoning modern frameworks. Server side rendering or static generation produces the same interface with meaningful content in the initial response, and it is faster for humans too, which is the usual pattern in this area.

Write it once, covering the category question, the problem question, the comparison question, the competitor question and the branded question. Fifty is a workable minimum. Then freeze it, and if you must add prompts later, add them as a separate cohort so the original series stays comparable.

Then Measure Again, and Keep Measuring A single snapshot tells you very little. Assistants vary their answers between sessions, between accounts and between model versions, so one run is a sample and not a verdict. Re-run the same prompt set on a fixed schedule and watch the trend rather than any individual answer.

Getting onto that list is not luck and it is not a trick. It is a sequence of fairly unglamorous steps that make it easy for a model to find you, understand you and feel safe naming you. This is what that sequence looks like in practice. generative engine optimization

Make Sure the Crawlers Can Actually Read You A surprising number of brands are invisible for the dullest possible reason. Their robots.txt blocks the crawlers that feed AI systems, or their content only appears after JavaScript executes, or their key pages sit behind a form.

Run each one across the assistants your customers use, and write down the answers verbatim. Do this from a signed out session so your own history does not colour the result. What you want at the end is a simple table: which prompts named you, which named competitors, and which sources got cited.

The complication is that AI systems use several distinct agents for different purposes. One may crawl for training corpora, another may fetch pages live when composing an answer, and a search provider's traditional crawler may feed both search results and an AI summary.

Check which agents you allow, confirm your important pages render meaningful content without scripts, and make sure nothing critical is trapped in a PDF or an image. This is the cheapest work in the whole discipline and it is routinely skipped.

What llms.txt Proposes It is a proposed convention: a file at your root offering a curated, plain text guide to your site for language model consumers, pointing at the documents you consider authoritative.

Marketing copy does not get quoted. A paragraph of adjectives about your commitment to excellence contains nothing a model can attribute, so it is skipped in favour of a competitor who wrote a plain answer. Write the plain answer. generative engine optimization

Run Each Prompt Multiple Times Generation involves randomness and retrieval can return different pages between runs, so a single answer is a sample. Three runs per prompt is the practical minimum and five is better where the stakes are high.

Control the Session Conditions Personalisation quietly corrupts this. Run from a signed out session, or a fresh session with memory and history disabled, and do not use an account that has been researching your own company all week.