How Do You Audit What AI Engines Say About a Competitor?

Why would you audit what AI engines say about a competitor?

Because the answer tells you which sources are shaping your category, and those sources are addressable. When an engine describes a rival confidently, it is drawing on specific pages written by specific people. Finding those pages is competitive research you can act on. Reading the answer itself is not.

Most people run one query, see a competitor named, feel bad, and close the tab. That is not an audit. That is a mood. The useful version produces a list of URLs and a list of claims, and it takes about two hours.

I have written more than three hundred and fifty articles on AI answer engines, schema and E-E-A-T, and this is the exercise I run before I touch anybody's content strategy. Not because the answers are truth, but because they are a readable summary of what the open web currently says about a market.

What are you actually looking for?

Three things. The claims an engine repeats about the competitor, the sources it cites when asked, and the gaps where it hedges or says it does not know. The third is the most valuable and the one everybody skips, because a hedge marks a question your category has not answered in public.

Claims tell you the positioning that has stuck. If three engines all describe a rival as the option for enterprise compliance, that phrasing came from somewhere and it is now doing work on their behalf in conversations you are not part of.

Sources tell you where to compete. If the answer leans on a review site, a documentation page or a single well written comparison article, those are the surfaces that matter, and some of them are surfaces you can appear on too.

How do you build a prompt set worth repeating?

Write ten to fifteen questions a real buyer would ask, phrase them the way a buyer would, and freeze them. Mix category questions, direct comparisons, objection questions and the awkward ones. Then never edit the list mid audit, because changing the question changes the answer and you lose the ability to compare.

The category questions look like what tools do teams use for a given job. The comparisons name two vendors. The objection questions are the ones your sales calls actually contain: is it expensive, is it hard to migrate to, does it handle a particular case. Those last ones are where engines reveal what the market believes.

Keep a handful that do not mention any vendor at all. Unprompted mentions are worth more than prompted ones, because they tell you who the engine reaches for when nobody has put a name in front of it.

Why does one run tell you almost nothing?

Because these systems are not deterministic. Ask the same question twice and you can get different wording, different emphasis and a different set of cited sources. Anything you conclude from a single answer is an anecdote, and it will not survive the next person who checks it.

So run each prompt several times, across more than one engine, and record every run rather than the one you liked. What you are looking for is what repeats. A source that appears in one run out of six is noise. A source that appears in five out of six is a fact about your category.

Logged in accounts add another layer, because history and personalisation can shape what you see. If you want results that describe the market rather than your own browsing, run them clean. I treat that the same way I treat any measurement problem, which is why I am insistent about building a baseline before you change anything.

What should you record for each run?

Four fields and nothing else: the exact prompt, the engine, the answer verbatim, and every source it cited. Add the date. Resist the urge to summarise while collecting, because your summary will quietly become the evidence and you will lose the wording the engine actually used.

The verbatim answer matters because phrasing is the finding. There is a real difference between a rival being described as built for enterprise teams and being described as popular with enterprise teams. One is a positioning claim, the other is a social proof claim, and they call for different responses.

A spreadsheet is enough for this. I keep one row per run in Airtable or a Google Sheet, because the analysis step is counting, and counting is easier when every run is a row rather than a paragraph in a document.

How do you turn the log into something you can act on?

Count the sources. Sort by how often each domain appears across all runs and engines. That ranked list is the deliverable, and it will usually be shorter and stranger than you expect. Then read the top five pages properly and ask what each one does that your equivalent page does not.

Often the answer is unglamorous. The page is older. It answers the question in the first paragraph instead of the fifth. It names specifics where yours generalises. It has a table of plain facts that a machine can lift without interpretation.

Separately, count the claims. Which statements about the competitor show up again and again. Those are the things you will be compared against whether or not you address them, so decide deliberately which ones you want to answer on your own pages and which you want to leave alone.

What you should not do is write a page arguing with the engine. Answers change, and a page built to rebut one is dated the moment it does. I made this case at length in why AI answer engines cite competitors instead of me: the durable move is to become a better source, not a louder one.

What does it mean when the engine gets your competitor wrong?

Usually that the record is thin, stale or contradictory, not that anyone is lying. A wrong answer about a rival is a warning about your own exposure rather than an opportunity, because whatever mechanism produced it will produce the same thing about you when your public record has the same gaps.

It is also a bad idea to build marketing on it. Pointing out that an AI assistant said something unflattering about a competitor is a claim with a short shelf life and a reputational cost, and it invites the obvious retaliation.

The useful reading is diagnostic. If engines get a rival's pricing wrong, go and look at how pricing is published on their site and on yours. The difference between a page that gets summarised accurately and one that does not is usually structural and boring, and it is fixable in an afternoon.

How often is this worth repeating?

Quarterly for most businesses, monthly if your category is moving fast or you have just changed positioning. The value comes from the comparison between runs, not from any single audit, and that only exists if the prompt set stayed frozen.

The second audit is where the method starts paying. A source that has climbed the ranked list is a competitor's content strategy working in public. A claim that has appeared since last quarter is a message that has landed. Neither is visible from a single snapshot.

Keep the old logs. When someone asks whether the work you did six months ago changed anything, a frozen prompt set run twice is the closest thing to an honest answer you will have, and it is considerably better than a feeling.

What should you do next?

Write your fifteen questions today and run them five times each across two engines. Log the prompt, the engine, the answer and the sources. Count the domains. Read the top five pages. That is the whole audit, and the ranked source list is the part you will still be using in six months.

Then pick one gap where engines hedge and write the page that answers it properly. A question your category has not answered in public is the cheapest opening you will find, and it is visible to anyone willing to do the counting.

If you want this run properly on your category, with the log and the ranked source list handed over rather than a slide about AI visibility, reach out. I work on fixed fees and I am happy to tell you what I would do first. Let's chat.