The default sentence

Start with the mechanic, because the mechanic is the whole argument. A buyer opens an assistant and asks who they should talk to about a problem you solve. The model does not open your website and read your positioning statement back to them. It produces its own description of you — a sentence, maybe two — assembled from what it has absorbed and can corroborate. Then it does the same for two or three other companies, ranks them loosely, and hands over a shortlist.

That sentence is now the first thing most buyers learn about your company. It has no author, no approval chain and no archive. It is generated fresh, per buyer, per phrasing of the question, and then it is gone. You will never see the one that cost you the deal.

Why the default sentence is expensive

Because it is not wrong. If it were wrong, you would find out. Somebody would forward it to you with three exclamation marks and your legal team would have a view. Errors get corrected. Blandness does not — it just quietly removes you from consideration, and nobody involved experiences it as a decision.

And because it is shared. "A leading provider of AI-powered solutions for enterprise" describes you and it describes the two companies you are most often confused with. When three descriptions are interchangeable, the buyer has no basis to prefer one, so they fall back on the tiebreakers available to them: price, logo count, and whoever the model happened to name first. Every one of those is a worse outcome than the one your product deserves.

The compounding is the part that should worry an executive most. These systems weight consensus, and consensus weights what already exists. A company that becomes the answer gets cited, which makes it more likely to be the answer, which produces more citation. It is a positive feedback loop with a narrow entry window, and it is being resolved now, in a market where almost no one is competing for it deliberately.

Two industries are solving the wrong problem

The brand agency's answer is a better claim. Sharper language, a new tagline, a workshop, a deck. It is sometimes exactly right and it is never sufficient, because the mechanism at work is not persuasion. A model is not moved by how confidently you assert something. Assertion is the cheapest signal on the internet and every system that reads at scale has learned to discount it.

The GEO or AEO vendor's answer is a technical fix. Structured data, an llms.txt file, FAQ blocks, a tidy sitemap. Also worth doing — we do it — and also not the answer. Ahrefs looked at 137,210 domains and found that 97% of llms.txt files received no requests at all; AI bots never went looking for the ones that did not exist.¹ The tactic that has dominated two years of AI-SEO advice is, on the available evidence, largely unread.

Both industries are answering a question about expression. The question is about evidence. What a model treats as true about you is not what you say; it is what can be corroborated by sources that are not you. Your own site is a witness statement. The verdict is written somewhere else.

You cannot write the sentence. You can make it the cheapest thing for a model to say.

What actually moves the answer

Three things, and they are not equally weighted.

First, specificity. A model reaches for the generic description because you have not made a specific one available. "Autonomous inspection for offshore wind, certified to DNV standards" cannot be confused with a competitor. "AI-powered industrial intelligence" is not a claim; it is a category with your logo on it. Specificity is not a copywriting preference here. It is what makes a sentence retrievable at all.

Second, corroboration. One source saying a thing is an assertion. Several independent sources saying the same thing is, functionally, a fact — and that is the state you are trying to reach. We describe this as corroboration distance: the number of independent, model-visible sources that repeat the claim you need repeated. Zero is an assertion. One is a press release. Three or more is what gets restated back to a buyer. Closing that distance is slow, unglamorous work that no software product performs for you, which is precisely why it is defensible.

Third, legibility. The machine has to be able to read the page. This is the part that is genuinely technical, and the failure mode is more common than the industry admits. Vercel instrumented the major AI crawlers and found that GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot and PerplexityBot did not execute JavaScript.² Googlebot renders; most AI retrieval does not. A company whose site assembles itself in the browser is being read as a blank page by exactly the systems it most wants to be read by, and nothing in its analytics will ever tell it so.

The order matters more than the list

Legibility first, because nothing else can be received. Specificity second, because a legible page with nothing distinctive on it just delivers the default sentence in structured form. Corroboration last and longest, because it is the only one that compounds and the only one that cannot be bought in a quarter.

Run the order backwards — buy links and mentions before deciding what the claim is — and you spread the default sentence more efficiently. This is the single most common way money is wasted in this category, and it is being wasted at scale right now.

The uncomfortable part

Sometimes the reason a claim will not stick is that it is not yet true. Corroboration works because it requires other people to say the thing, and other people will not say a thing that is not so. When we find that, we say it, and the engagement changes shape: the work becomes making the claim true before making it visible. Firms that skip this step are selling amplification of a message the market has already declined to repeat.

We would rather lose the engagement than run that play. Not out of virtue. Because it does not work, and because in a market this full of manufactured proof, being a firm that says the inconvenient thing is itself the position.

What this means for a leadership team

Three practical consequences, in the order we would act on them.

Read your sentence. Not a share-of-voice dashboard — the actual sentence, across the systems your buyers use, with the sources traced. It takes days, not months, and it is usually the most uncomfortable document a marketing team has read in a year. Most companies have never done it.

Decide what should replace it, and hold that decision to a hard test: is this true, is it specific, and is it not equally true of the two companies we are confused with? Most positioning statements fail the third test and no one notices, because no one has ever had to see them side by side in a machine's output.

Then build the evidence, starting with the things only you can produce: original research, published method, real numbers, named people who will stand behind a claim. In a market where engagement is purchasable and case studies are decorative, verifiable evidence is the one asset that cannot be counterfeited — which is exactly why it is worth building.

Nothing here requires believing that AI will replace search, or that models are intelligent, or any other claim that will look silly in three years. It requires only one observation, which anyone can verify in ninety seconds with a phone: something is describing your company to your buyers, constantly, and you have not read what it says.