Competitive analysis for AI search

Almost every number in an AI visibility tool is a comparison. Share of voice is your mentions as a proportion of everybody’s. A leaderboard is a ranking within a set. “You were not named and three competitors were” is a statement about which three. Change the set and every one of those numbers changes, without anything having happened to your brand at all.

That makes the competitive set the single highest-leverage input in this kind of measurement, and it is almost always the least examined. It gets configured once during setup, by whoever was in the room, and is never looked at again.

What a competitor is, for this purpose

Not who you lose deals to. Who the engines name instead of you.

Those sets overlap and are not the same, and the difference is the interesting part. Your sales team knows the three companies that turn up in procurement. An assistant asked “what is the best tool for X” is composing an answer from what it has read, and what it has read is dominated by listicles, forum threads and directories. Those sources have their own idea of your category, assembled by people who were writing for traffic rather than accuracy.

So the set that matters is the one that actually appears in the answers. A company you have never lost a deal to, which is named in nine of the ten answers you want to win, is your competitor in the only sense this measurement recognises.

Building the set

Three sources, in descending order of usefulness.

• The answers themselves. Run your prompts, read who is named, and take the recurring names. This is the only source that is measured rather than assumed, and it routinely surfaces a company nobody internally had heard of.
• The pages the engines cite. Ranked listicles are explicit competitive sets, written down. If the same eight names sit on four of the pages models read about your category, that is the category as the machines understand it.
• Your own belief. Worth recording, and worth treating as a hypothesis. The gap between the list you would write from memory and the list the answers produce is itself a finding.

Size matters more than people expect. Too few and share of voice is meaningless — with two rivals, one of them having a good week moves your number by a third. Too many and the average drags toward zero as you add companies that were never going to be named. Somewhere between five and ten is usually where the number starts behaving like a measurement.

The failure this guide exists to prevent

Here is a real one, seen on a competing product. The project was configured for a founder social network. The dashboard reported the brand as “YouArgue” — a different company entirely — and tracked nine debate apps as its competitors. Every panel was populated. The visibility score was 0%, presented with complete confidence, on a trial with three days left.

Nothing about that dashboard looked broken. It had numbers, charts and a competitor leaderboard. It was measuring a company that was not the customer against rivals that were not in the category, and the only way to notice was to read the brand name at the top and recognise it as wrong.

This is what makes a stale competitive set worse than an empty one. A product with no data shows you it has no data. A product with the wrong set shows you a complete, plausible, confidently wrong picture, and every decision taken from it is taken in good faith.

Reading the comparison honestly

Once the set is right, three cautions about what the numbers mean.

A mention is not an endorsement. Being listed fourth in a paragraph explaining why the first three are better is a mention. If your tool cannot distinguish being recommended from being named, its share of voice is counting your losses as wins. Sentiment on a mention is not decoration for this reason: being described badly is worse than not being described.

Position within an answer is real but soft. Earlier is better, because readers stop reading. But an assistant’s ordering is not a ranking in the search sense — ask twice and it can differ. Treat a one-place move as noise and a sustained change as signal.

Compare like with like. Your competitor’s score and yours have to be computed the same way, from the same answers, with the same rules about what counts. A tool that measures your appearances carefully and approximates everyone else’s is not producing a comparison, and the direction of that error is always flattering.

Keeping it current

Categories move. A company launches, another pivots, a third gets acquired and stops being marketed. The set that was right in January quietly stops describing the market, and because every number keeps rendering, nothing announces the change.

The cheap discipline is a standing review. Once a quarter, look at the brands the answers named that you are not tracking, and ask whether any of them belong in the set. That list is free — the answers were captured anyway — and it is the earliest warning that your category has changed shape.

The signal to watch for is a name you keep skimming past. A company appearing in a quarter of your answers that nobody has decided about is either a competitor you have not admitted to or a false positive worth excluding deliberately. Both are better than leaving it undecided.

Where this connects

A competitive set is only as good as the questions it is measured against. If your prompts do not match how buyers actually ask, you are comparing brands on questions nobody poses. And if the engines are unsure which company you are, the comparison is polluted before it starts: mentions of a similarly named business will be counted as yours, or yours as theirs.