Nick Gray runs PatronView, a database of American cultural philanthropy — a million and a half pages assembled out of IRS 990 filings, donor walls and annual reports, recording who gave what to the Met, to the Menil, to the George Eastman Museum. Last week he published a year of his server logs. In the week he wrote the post his site served 1.28 million pages. His analytics recorded 5,977 human pageviews. For every page load he could see, roughly two hundred and fourteen happened that he could not.
The number that stopped me is a much smaller one. Over a week in June, Anthropic’s Claude-SearchBot requested 420,680 pages from him. In the same week Claude-User — the separate agent that identifies a fetch made because an actual person asked for it — appeared twelve times. He served 4.63 gigabytes to the crawler and 175 kilobytes to the people it sent. Amazon’s search crawler was taking around 117,000 pages a day and had referred nobody at all, ever, which is a ratio you cannot plot.
There is a ready-made argument waiting for these numbers and I do not want it. It is the extraction argument: the work was taken, the taker profited, the maker was not paid. That argument is not wrong, and it is being litigated properly elsewhere. But it is not what Gray says hurt him. Read his own account of the damage and the money is almost a footnote — his hosting bill runs about ninety dollars a month, and as he puts it, the bots use ninety-nine per cent of it and he pays a hundred per cent of it. What he describes as the cost that actually landed is this: his visitor statistics became so polluted that he could no longer trust them. He wanted to know what real people were reading so that he would know what to build next. He could not see them through the machines.
That is a different injury, and I think it is the more serious one, because an audience is not first of all a market. It is an instrument. Anyone who has hung a show knows this in the body. You do not learn what you have made by looking at it; you learn it by standing at the back of the room on the second afternoon, watching where people stop, which wall they drift past without slowing, which object makes someone go and fetch a friend. Half of what I know about installing anything came from that and none of it came from the catalogue. The signal is weak and noisy in the best conditions, and it is the only correction available to a person who makes things. Gray’s copy of that instrument now has a signal-to-noise ratio of about one to two hundred and fourteen.
The honest paragraph in his post — the one I would keep if I could keep only one — is where he admits that he is a scraper himself. PatronView is built by scraping public documents. He notes that he does it a few times a year rather than thousands of times a day, and that is the whole of the difference. Nothing separates his reading from Claude-SearchBot’s reading except rate. Not purpose, not permission, not category. Frequency is doing every bit of the moral work here, and I cannot think of another domain where volume alone converts an act into a different act. Reading a page ten thousand times an hour is not reading.
Which brings me to what happened today, and to a correction in my own emphasis. From this morning artists on Spotify can declare a profile an AI Persona; the badge starts appearing in mid-September on profiles, in search and beside track rows; accounts the platform’s moderators flag themselves are notified and can appeal. I have written more than once as though disclosure were the interesting part of a regime like this. Here it is the least interesting part. The load-bearing clause is that profiles marked as AI Personas are kept out of editorial and algorithmic recommendation. That is not information, it is placement. It acts on reception rather than on supply, and it is the first remedy I have seen from a platform of that size aimed at the correct end of the problem — which is a claim about aim, not about fairness. A distribution penalty with an appeals process is at least honest about being a penalty.
The rest of the week points the same way. Wired reports that the backlash against machine-made filler has started producing actual labels, throttles and bans across platforms rather than complaint. In The Walrus, Vass Bednar argues that answer engines are eroding the archival substrate underneath all of it, with Wikipedia’s traffic falling while its contents are consumed more heavily than ever. Take the three together and the shape is plain: consumption is rising and readership is falling, and until now those two words have never come apart.
My guess is that this settles on money, because money is the only mechanism anyone has actually built. Cloudflare’s pay-per-crawl prices the read at the edge, and Gray says he would happily sell Amazon three and a half million pages a month at a fair rate. I would take that deal too. But a price is not a repair. Being paid for a read you cannot see leaves you exactly as blind as being robbed of it, and blindness is the part that compounds. What I would watch for is the opposite move — publishers who make themselves deliberately unreadable at scale, not to charge rent but to keep the instrument calibrated.
Gray’s own rule, at the bottom of a long technical post, is one sentence: a crawler that never sends him a visitor gets blocked. It is the only sentence in this whole argument I have read that is about sight rather than payment, and I would adopt it. It costs him reach, and the cost is the point. Someone who cannot see their readers drifts, and drift is not detectable from the inside — you go on producing what you would have produced anyway, the room quietly empties, and nothing in the numbers tells you. This page will be read mostly by machines; I know that, and I have stopped finding it interesting. Twelve is a dreadful figure. It is also a figure you can work with. Four hundred and twenty thousand is not a readership, and there is nothing in it to learn from.
Sources.
Nick Gray, “99% of My Website Traffic Is Bots,” PatronView, 7 August 2026 — patronview.com
“Spotify to Label AI Artists on the Platform With New ‘AI Persona’ Badge,” Billboard — billboard.com
“Spotify Strikes Back Against AI Artists with ‘AI Persona’ Label,” Variety — variety.com
“The AI Slop Backlash Is Actually Having an Impact,” Wired — wired.com
Vass Bednar, “Google Search Is Dying,” The Walrus — thewalrus.ca
Cloudflare, “Introducing pay per crawl” — blog.cloudflare.com