When other people talk about you

For a long time, the answer to how a business gets found online was a list: pages, structure, load time, links. All things you do on your own website. All things you control.

That answer is now half the truth, and the missing half is uncomfortable.

Two questions, two entirely different routes

There are two kinds of question people put to an AI, and they get answered from entirely different sources.

The first kind is checkable: when is the business open? Do they do emergency call-outs? Where exactly are they? Roughly what does it cost? For questions like these, an AI reaches preferentially for the source the business itself owns. An analysis of 6.8 million AI citations found 44 per cent came from first-party sources — the brand's own website. Count in directories a business can also control and it reaches 86 per cent. I covered those figures in Website or social media, and they still stand.

The second kind is a judgement: who is best? Who would you recommend? Which supplier suits my case? Here everything changes.

What the numbers say about the second kind

Ahrefs studied 75,000 brands, measuring which signals go together with visibility in AI answers. The result is uncomfortable for anyone who has spent twenty years thinking about links:

Web mentions correlate at 0.664. Backlinks only at 0.218.

Louise Linehan, who led the analysis, draws a clear conclusion: “Online brand mentions appear to be more impactful than link building.”

The backlink — for years the currency of search optimisation, traded, bought, managed in spreadsheets — is by comparison the weaker signal. What counts is the mention. Even without a link. Even without anyone pointing at you.

In a follow-up analysis, the single strongest correlation of all was mentions on a large video platform: around 0.737. And a study of over a million AI-cited links found that roughly 82 per cent of citations come from third-party editorial sources — what the trade calls earned media.

Why this does not contradict the 44 per cent

At first glance the two sets of figures contradict each other. 44 per cent from your own site — and at the same time 82 per cent from other people's. Both cannot be true.

They can. They measure different questions.

Faced with a question of fact, the AI looks where facts live: with you. Faced with a question of judgement, it does what a sensible person would — it does not ask the seller, it asks around.

That is not a technical quirk. It is a rather reasonable rule, and it describes exactly what all of us do when looking for a tradesperson: we read the opening hours on their site. Whether they are any good, we ask someone else.

The part you cannot build

Here it gets uncomfortable, and I will say it as I mean it: the second part cannot be built.

I sell digital foundations. I can set up a website so that the answer to every question of fact is cleanly available — machine-readable, current, free of contradictions. That is craft, it is plannable, it can be ordered.

Mentions cannot be ordered. They can be earned, and they can be faked, and the second gets found out.

What you can do is narrower — and it is still plenty:

Have something worth talking about. That sounds like a motivational poster, but it is the hardest requirement of the lot. A business doing the same thing as twenty others and describing it the same way gives nobody a reason to name it.

Publish your own data. This is the most underrated lever there is. Anyone who measures something nobody else measures and makes it openly available gets cited — not out of kindness, but because it is not written down anywhere else. Every trade has numbers like these. Most businesses do not know they are sitting on them.

Be reachable for the people who write. Trade publications, association media, local business desks are constantly looking for someone who can say something concrete, and rarely find anyone who replies. Turn up once and you tend to get asked again.

Be visible where technical things are read. This is where video and professional networks come in — not as an end in themselves, but because that is where the background noise forms that gets measured.

Our own figure — and what it does not mean

We measure this on ourselves, with the same questions and the same models: with web search every week, without web search once a month. As of 7 September: in web search, btlabs is named in 14 of 30 queries. In pure model knowledge — no access to the web — in zero of 75.

That zero is not a puzzle, it is arithmetic. Our first article appeared on 24 June, which was 75 days before that measurement. The models we query were trained on data gathered long before that. There simply was no route by which we could have got in.

This is why we keep the two figures apart. Anyone who merges them is, on a young domain, mostly measuring its age. And anyone concluding from such a zero that mentions do not work is concluding too early: the channel that reacts quickly is showing 47 per cent right now.

For you it means the same as for us. The mentions that arise today do not decide today's answer, but the one two years from now. That is uncomfortable, because there is no lever you can pull faster. But it is also the only reason to start now rather than later.

Where mentions actually come from

“Get mentioned” is about as useful a piece of advice as “be successful”. So here are the channels, sorted by what they cost — not in money, in attention.

Trade associations and guilds. The most underrated channel, because it looks boring. Member directories, technical pieces in association media, reference lists: all editorial, all lasting, all without an advertising budget. Anyone organised in their trade already has this channel and mostly does not use it.

Local business desks. They constantly need someone who can say something concrete, and rarely find anyone who answers. Being reachable once often means being asked again. It costs an email and the willingness to give information even when there is nothing to sell.

Trade forums and professional communities. Where colleagues and customers ask technical questions. A useful answer under your own name is a mention that stays — and it gets read by the systems that later make recommendations.

Video. The analysis names mentions on a large video platform as the strongest single signal. That does not mean every business must produce videos. It means: if your subject is discussed there and your name comes up, it weighs more than elsewhere.

Reviews and map services. No substitute for editorial mention, but they add to recognisability — and it is the channel almost every business already has.

What stands out going through the list: not one of these channels is new. They are the same ones a business became known through twenty years ago. What has changed is only who is reading along.

One year, planned realistically

The most common mistake at this point is a plan that stalls after three weeks. So here is an order of magnitude a business with other work can actually sustain.

The first three months: tidy up and have one thing. Get the facts side in order — that is the precondition without which everything else runs into the void. In parallel, answer a single question: what do we know or measure that others do not? Not the marketing answer, the honest one.

Months four to six: put that one thing out. Your own figures, your own experience, your own case — in a form that can be cited. Dated, with the method attached, openly accessible. A single piece like that is worth more than twelve general ones.

Months seven to nine: become reachable. Let the association editors, the trade press, the local business desk know there is someone who answers on this subject. That is a short message, not a campaign.

Months ten to twelve: check whether anything happened. And on the right number — not visitor counts, but whether an AI names the business by itself when asked about the service without the brand name.

That is little for a year. It is more than most do in five.

And a word on expectations: it is quite likely that after twelve months, in the models that answer without searching the web, there will still be nothing. That is not failure. Those systems are retrained at intervals; what is created today can arrive in the generation after next. Anyone who cannot sit with that should not start the second part at all — and should then at least do the first part properly.

The question nobody asks

People who talk about mentions almost always talk about volume. How many, how often, how quickly more.

The more interesting question is a different one: in what context does the name appear?

A language model does not merely register that a business exists. It registers which words it appears beside. Someone mentioned ten times in contexts that have nothing to do with their competence — sponsor lists, association newsletters, address directories — has ten mentions and no profile.

Someone mentioned three times where their field is written about has three mentions and a placement.

That is the practical consequence I draw from the numbers, and it is written that way in none of the studies: mentions are not a counter, they are a context. Which is why one technical contribution is worth more than ten appearances in a list — and why the question "where is my subject written about" matters more than "how do I turn up more often".

For a small business that says something reassuring: this is not about reach you cannot afford. It is about a few places where the name sits in the right surroundings.

What I am expressly not claiming

• I am not claiming backlinks have become worthless. They correlate more weakly, not negatively. Anyone who has them loses nothing.
• I am not claiming correlation is cause. That mentioned brands are more visible may simply mean well-known brands are both. The studies measure an association, not a causal chain.
• I am not claiming you should buy PR. Paid mentions are recognisable as such, increasingly so. What gets measured is editorial mention.
• And I am not claiming the website becomes unimportant. It still answers every question of fact. Without it the best reputation is useless, because the concrete details an AI would need to recommend you are missing.

What that means for the order of things

Anyone starting both at once usually does both by halves. So here is a simple rule of thumb: the foundation is a project, mentions are a habit.

A project has a beginning, an end and someone responsible. Getting the facts side in order is exactly that — and once it is done, it largely stays done.

A habit has no end. Being reachable for enquiries, publishing something of your own now and then, staying visible in your professional surroundings: that cannot be completed, only practised. Which is why it does not belong in a project plan but in a calendar.

A last note on the numbers

Every figure quoted here has a date and a source, and none of them says what will happen. They describe how the systems behave now.

I mention it because this field carries a lot of prediction and rather little measurement. A prediction cannot be checked; a measurement can — and if it says something different in a year, this piece will need rewriting. That seems to me the more honest deal to offer a reader.

What would tell me I am wrong

Two things would change my mind.

First, if the systems giving answers moved back to citing mostly supplier claims for judgement questions too. That is technically possible and would invert the whole argument.

Second, if it turned out the measured correlations do not hold for small local businesses. All the studies cited look mostly at larger brands. Whether the same associations apply to a three-person trades business is open — I think it plausible, but it is not proven.

What to take away

Your website supplies the facts about you. Whether an AI recommends you is decided by others.

The first is craft and lies entirely in your hands. The second is slow, indirect and cannot be ordered — and precisely for that reason it is the part worth starting earlier than it feels like.

How the technical side works is in From SEO to GEO. What can actually be measured about your visibility, in What you measure when position one is worth nothing.

Sources

• Correlation analysis across 75,000 brands: web mentions 0.664 against backlinks 0.218; in the follow-up, mentions on a large video platform around 0.737. Ahrefs, 26 May 2025.
• Study of over a million AI-cited links: around 82 per cent from earned media. Muck Rack, 2025/2026.
• Analysis of 6.8 million AI citations: 44 per cent first-party, 86 per cent from brand-controllable sources. study by a location-data provider, 2025.