Somebody just asked an AI where to buy what you sell

Right now, somewhere, a person is typing “best pool pump for a small yard, delivered in Cape Town” into a chat window instead of a search box. They will get a straight answer. It will name two or three businesses. And it will be assembled, in part, out of pages belonging to companies who will never know the conversation happened — no impression, no click, no line in anyone’s analytics.

We wanted to know how often that happens. Anthropic publishes the raw usage data behind its Economic Index, so we downloaded it and went looking.

One in seven

Across every conversation people had with Claude in May 2026, 14.55% were somebody trying to find something out — research and retrieval, up from 13.78% a month earlier.

That makes looking things up the fourth most common thing anyone does with an AI, ahead of writing software. Which is the part worth sitting with: over that same month, research went up and coding went down.

It doesn’t behave like a work tool

The dataset splits that into two categories, and the smaller of the two has a fingerprint worth describing. It is two-thirds personal, not work. Its output is the answer itself nearly three-quarters of the time — not a document, not a report, not code. And it is the shortest task in the entire dataset: about thirty-seven minutes of human work saved, against more than eight hours for a software task.

Personal. Quick. The answer is the whole point. There is a name for a tool people use like that, and it isn’t “assistant”. People have taken the thing they used to do in a search box and moved it, more or less intact, into a conversation.

And they’re asking about things they intend to buy

This is where it stops being an interesting statistic and starts being your problem. Underneath those categories sit the individual topics: buying and investing, finding the best price, destination research, gift buying, product compatibility, fashion shopping, secondhand.

Every single one of them grew between April and May. Not most of them — all eleven we tracked. Over the same period, software development fell.

None of these are enormous on their own; buying and investing is 2% of all conversations. But they are not abstract research tasks. Each one is somebody deciding where to spend money, and each one now resolves inside a conversation where a few businesses get named and everybody else simply does not exist.

There is no page two to be on. There is no tenth blue link. You are the answer, or you are absent.

Nobody’s coming to check

The obvious objection is that this is fine, because surely the person clicks through afterwards. They don’t. In early 2026, 68% of Google searches ended without a click of any kind — up from 60% two years earlier, on a measurement that excludes the Google mobile app and is therefore, if anything, generous. When an AI summary sits at the top of the results, Pew found people clicked a normal link 8% of the time instead of 15%.

And the machines doing the reading are not sending anyone back either. Cloudflare measured this directly: in July 2025, Anthropic’s crawlers read 38,000 pages for every single visitor they referred. OpenAI’s ratio was around a thousand to one. Google’s, for comparison, was five to one. Roughly 80% of that crawling wasn’t even for answering live questions — it was for training. Those Cloudflare figures are a year old now, and they’re the most recent we could verify from the source rather than from somebody’s blog. Take the exact numbers loosely; the shape is not really in dispute.

Your content is being read at industrial scale, turned into answers, and handed to people who never arrive. The only version of that trade where you get anything back is the one where the answer names you.

What we can’t tell you

We should be straight about the limits, because most people writing about this aren’t. This is Claude — one assistant, with users who skew technical and paying. It is not the market, and ChatGPT is far larger. More importantly, this data shows a destination, not a migration: it proves a great many people ask an AI to find things out, but it contains no evidence at all that those people stopped using Google to do it. Only two months are comparable, so the trends here are a direction, not a curve. And the “thirty-seven minutes” is a model’s estimate of human effort, useful for comparing categories against each other, not a stopwatch.

What survives all of that is narrow and, we think, solid: a large and growing share of AI use is people looking things up, the part of it with money attached is rising every month we can see, and the click that used to follow has mostly stopped happening.

The awkward part

For twenty years the job was to rank. The job now also includes being readable — by a machine that will summarise you for someone you’ll never meet.

Those are related jobs, not the same job. Good SEO is still the foundation, and nothing here changes that. But a site can rank perfectly well and still be useless as a source, because most of what makes it a good source is invisible to human visitors. Whether an engine can reach the pages at all. Whether the content survives being read without JavaScript. Whether the business is identifiable as a real entity rather than a logo. Whether anything on the site is shaped like an answer to a question somebody would actually ask. Whether any of it exists in the language the question was asked in.

None of that is hard. It’s just nobody’s job yet. We know, because we’ve been counting. Across the first 62 websites that ran a LoopCite scan, the median readiness score was 62. Nothing scored above 78. Not one site hit all five targets — the best any of them managed was three.

So: the questions have moved, the click is already gone, and the sites that would need to be the source aren’t ready to be one.

One thing no one can sell you: a guarantee that an AI will cite you. Engines decide what to quote and no vendor controls that. What you can control is whether you’re eligible to be quoted at all.