Which engine actually sends local traffic in 2026?

Every share-of-answers chart you have seen is somebody’s guess dressed as data. The number that matters is not the market’s — it is yours.

The honest answer first

Nobody outside the model providers knows how local answer traffic splits between them, and the providers do not publish it. Every pie chart claiming otherwise — including an earlier version of this page — is an estimate wearing the clothes of a measurement.

So this post does not have that number, and we are not going to invent one.

If a vendor shows you a share-of-answers chart, ask what they counted. The useful answer is a method. The usual answer is a vibe.

What the question usually means

When an owner asks which engine matters, they rarely want a market-research figure. They want to know where to put a limited amount of effort — whether to worry about one assistant in particular, and whether the work is different for each.

That version has an answer, and it is a much more useful one: the effort is almost entirely shared. Nearly everything that gets you named in one place is the same work that gets you named in the others, because the engines are reading the same material about you. There is no separate campaign to run per assistant, and a vendor offering you one is selling the same job several times.

Why nobody can tell you

It is worth being specific about why the number is missing, because the reasons are structural rather than temporary. Nobody is about to release it next quarter.

The providers hold the only complete record and treat it as commercially sensitive. What they publish is at the level of total users, occasionally, in a blog post, with no breakdown by intent — and local buying intent is a thin slice of everything an assistant is asked to do. Knowing how many people use a product tells you nothing about how many of them used it to choose a plumber.

The old proxies have gone quiet, too. Web traffic was countable because a visit carried a referrer: you could see which search engine sent it. An answer often ends the journey — somebody reads a paragraph, taps a phone number and calls you, and no analytics tool anywhere records that the sentence naming you came from one assistant rather than another. The most valuable outcomes are the least traceable ones, which is a miserable property for anyone trying to build a chart.

Panel data does not rescue it either. The panels that estimate assistant usage are built from browser extensions and volunteer devices, skew heavily toward desktop and toward technical users, and mostly cannot see inside mobile apps at all — which is where a person standing in a leaking kitchen actually asks. A panel can tell you something about who opens which app. It cannot tell you which app named a business in a town it has three respondents in.

What the charts are actually counting

When you do see a confident split, it is usually one of three things, none of them the thing the label claims.

It may be app usage: how many people opened each assistant, all questions included, from cookery to code. It may be traffic referred to publisher sites, which measures the assistants that link out most rather than the ones that answer most. Or it is a survey — people asked which assistant they use, answering from memory, in a category where memory is famously poor.

Every one of those is a real number about a real thing. None of them is the share of local buying answers, and the distance between them and that claim is where the chart stops being research and starts being decoration.

The boundary is blurry anyway

There is a further problem underneath the missing data, and it is the one that makes the question quietly unanswerable rather than merely unanswered.

Which engine assumes tidy product boundaries, and there are none. An AI answer now appears above ordinary search results, inside a maps app, inside a browser sidebar, inside a phone’s built-in assistant, and inside the chat window of a company whose model belongs to somebody else. Plenty of people read a generated answer about a local business every week without believing they used an assistant at all — they searched, the way they always have, and the top of the page had changed.

Ask those people which assistant they use and they will name whichever chat app they have opened deliberately. The answer that actually decided a call may have been served somewhere they would never think to mention, by a model they could not name. Any survey built on that self-report is measuring recognition of a brand rather than exposure to an answer.

So even a perfectly honest share number would need to say what it was counting as an engine before the percentage meant anything at all.

What can actually be measured

Your own coverage can. That is a smaller question than the market’s, and it is the only one that changes what you should do on Monday.

An audit asks each engine the same buying questions a customer would type, in your city, without ever naming your business — naming it manufactures the answer. What comes back is checkable: whether you were named, which sources the engine leaned on, and what it believed about you. Run that across the engines and the picture you get is yours, not an industry average that may not describe your trade or your town at all.

It is also repeatable, which the market number is not. Ask the same questions next month and the difference is attributable: either something you did moved, or something in the market did. That is a far more useful instrument than a share figure would be even if a true one existed, because a share figure tells you about an industry and this tells you about a business.

The questions, the engines, the number of runs and the way a mention is decided are all written down in full, for the plain reason that a method you cannot inspect is worth exactly as much as a chart you cannot source.

Why the split matters less than it sounds

The engines are not drawing on separate worlds. They read the same directories, the same review sites, the same community threads. A business described consistently across those sources tends to turn up wherever it is asked about; a business described three different ways tends not to.

That is why the ranking question is less useful than it looks. You do not have an engine problem, you have a source problem — and the sources are shared. Fix the description once and you have fixed it for all of them.

There is a version of this that does hold up, and it is worth separating from the version that does not. Engines differ in temperament: some quote a source almost verbatim, some summarise several, some hedge and name nobody. Those differences are real, observable in your own report, and occasionally worth acting on. What does not hold up is the leap from that to a national percentage — an engine that is generous with names in your trade and your town may be nothing of the sort two counties over.

We do weight the engines, and that is an assumption

Here is the part that would be easy to leave out.

A visibility score cannot treat every engine as equally important, because they are not. So the score weights them — and those weights are a judgement, not a measurement. They are the one place in this product where a number came from reasoning rather than from something we counted, and the honest thing is to say so on the page rather than bury it in a methodology footnote.

Three rules keep that from becoming the very thing this post objects to. The weights are published rather than hidden. They renormalise across whatever engines were actually probed, so an engine we did not ask cannot silently cost you points you had no way to earn — and an engine that was not probed is never drawn as a miss. And any change to them bumps the scoring version, so no score is ever compared against one produced under a different set.

That is the difference between an assumption and a fabrication: an assumption is stated, bounded, and versioned, and you can disagree with it out loud. Nothing about the weighting is presented to you as a fact about the market.

What would have to exist before we published a number

For completeness, because it is a fair question to put to anyone who refuses to answer one.

We would need a sample of real buying questions, from real locations, at a national scale rather than in the handful of cities where we happen to have customers. We would need it repeated over time, because a single week is weather rather than climate. We would need it segmented by category, since visibility in trades behaves nothing like visibility in restaurants. And we would need the outcome — whether the answer led to a call — which is the piece nobody currently has.

Until then, the responsible version of this page is the one you are reading. When the measurement exists, it will arrive with its method attached, and you will be able to argue with it.

So what should you actually do

Work in the order the sources dictate, not the order the logos suggest.

Start by finding out what the engines are reading about you and whether it agrees with itself — that is the sources that keep coming back, and it is the input every engine shares. Make those sources say the same thing. Then make your own site the place that answers the questions your buyers ask, because a site that states nothing cannot be quoted by anything.

Only after that is it worth wondering which assistant your customers prefer — and by then the answer will mostly have stopped mattering, because the work that wins one wins the rest. If you are new to all of this, start with what answer engine optimisation actually is, then read what a visibility score is made of before you let anybody sell you one.