We added a fourth AI engine to the index. The score did not move.

On 19 July we added Perplexity to the index. It is the only assistant we measure that is built search-first — it retrieves before it answers, rather than answering from what it already holds. The reasonable expectation was that adding it would move brands' scores, probably upward for anyone with a well-built site.

The next day we re-measured a brand across all four engines. The score came back 8 out of 100. Two days earlier, on three engines, it had been 8 out of 100.

That is the finding, and it is more interesting than a change would have been.

What actually got measured

The reading covers 252 individual answers, judged across ChatGPT, Gemini, Claude and Perplexity, against a fixed set of 37 questions phrased the way a buyer would phrase them — the same 37 that every brand in the category is measured against.

Perplexity named the brand in 10.8% of its answers. ChatGPT named it in 10.8% of its own. The two engines, built on opposite principles, produced an almost identical rate.

Where Perplexity differed was in recommendation. It was the least likely of the four to actively suggest the brand — roughly 1 in 74 answers, against about 1 in 37 for ChatGPT and Claude. The engine that looks things up was the most reluctant to endorse what it found.

Then what did change?

The mechanism. Every mention we record is classified by how it was produced: did the model already hold the brand, or did it go and look it up mid-answer?

Before Perplexity, 33% of this brand's mentions were retrieved and 67% came from what the models already knew. After, retrieved rose to 41%.

So the composition of the visibility changed materially while its quantity did not. The brand is being found by search more often than it was, and that is not translating into being named more often, or recommended more often.

That distinction is the whole of the practical advice. Mentions that come from memory are earned slowly, by being written about across the web over years. Mentions that come from retrieval are earned by being findable and quotable right now — and they are the half a business can actually move this quarter. A brand whose visibility is 41% retrieved has considerably more control over its own number than one sitting at 5%.

Why a single number could not have told you this

A visibility score that goes from 8 to 8 looks like nothing happened. On a dashboard, that is exactly what you would see: a flat line, and no way to tell whether the flatness means stability, stagnation, or two opposite movements cancelling out.

Here it was the third. Retrieval rose by eight points of composition and the outcome held. That is a real event, and it is invisible to any instrument that reports one figure.

It is also why we report a range rather than a point. The same reading puts the brand's mention rate at 10.7%, with the true value somewhere between 7.5% and 15.1%. That interval is wide despite 252 sampled answers, because assistants answer the same question differently every time. Any month-on-month movement inside it is dice. Most of what is sold as improvement in this category is movement inside an interval nobody has shown you.

What to take from it

Adding engines does not add visibility. Coverage tells you where you stand; it does not change where you stand. Anyone selling extra assistants as though each one were extra reach has the causation backwards.

Search-native does not mean easier to win. The engine that retrieves was the one least willing to recommend. Being found and being endorsed are different achievements, and the second is harder.

Look at the composition, not just the number. If most of your mentions are retrieved, your content and its citability are the lever. If almost none are, you have a slower and more expensive problem, and you should know that before you spend against it.

None of this is visible from measuring one brand once. It comes from asking a fixed set of questions, repeatedly, across several engines, and recording how each answer was produced — which is what an index does and a dashboard does not. The method is published.

Run a scan to see the readiness half of your own picture, free.