Published 7 October 2026.
Most posts about the future of search can’t be marked wrong. They say AI will “change everything” and never say by when or how you’d tell. This one makes three calls, each with a date and a test.
Here’s the present as I can measure it. When an answer engine uses someone’s work, three things can happen. It links to the source. It names the source without a link. Or it does neither. Only the first one can send a click. That’s what I mean by a smaller click: it isn’t gone, it’s one outcome out of three.
Publishers are already acting on this. On 2 October 2026, nine of the UK’s ten biggest newsbrands blocked at least one AI crawler, and training bots were blocked more often than user-triggered fetchers (my crawler study; sample and “blocked” defined in the method note). Small sites mostly aren’t deciding at all. Under RFC 9309, silence in robots.txt is a yes, so the platform default decides for them (the default vote). And when engines quote a price, they rarely say when it was true (undated prices).
What this post will not claim
- That AI search will replace Google.
- Anything about traffic or revenue, for publishers or anyone else.
- That a crawler block causes anything. A block shows what a site asked for, not what happened next.
- Anything outside the three tests below.
P3: Most answers won’t carry a link
- Claim: On the first run of my registered link, mention or nothing study, fewer than half of the scored answers will be “linked”.
- Called: 7 October 2026.
- Check date: 7 January 2027.
- Test: Run the study as registered: the same 30 UK news queries, one row per query and engine, each answer scored as linked (at least one clickable link to a news source), mention only (a news brand named, no link) or neither (no news source named).
- Right if: linked rows are fewer than half of all scored rows.
- Wrong if: linked rows are half or more of all scored rows, or the study hasn’t been run and published by the check date.
- Inconclusive if: an engine named in the study plan no longer exists on the run date, so the registered method can’t be run as written.
- Baseline: not measured yet. The study was registered on 2 October 2026 and hasn’t been run.
P4: Every engine will name without linking at least once
- Claim: On the same first run, every engine in the study will give at least one “mention only” answer.
- Called: 7 October 2026.
- Check date: 7 January 2027.
- Test: Count mention-only rows per engine in the P3 run.
- Right if: each engine has at least one mention-only row.
- Wrong if: any engine has zero mention-only rows, or the study hasn’t been run and published by the check date.
- Inconclusive if: an engine named in the study plan no longer exists on the run date.
- Baseline: not measured yet.
P5: The publisher split holds for a year
- Claim: A year after the first snapshot, at least nine of the same ten UK newsbrands will still block at least one of the same 15 AI crawlers.
- Called: 7 October 2026.
- Check date: 2 October 2027 (the monthly snapshot for that date).
- Test: Read robots.txt for the frozen sample (news.sky.com, theguardian.com, dailymail.com, independent.co.uk, bbc.co.uk, mirror.co.uk, uk.news.yahoo.com, thesun.co.uk, express.co.uk, telegraph.co.uk) and count sites with a whole-site
Disallow: /for at least one of the 15 crawlers, using the definitions in the method note. - Right if: 9 or 10 of the 10 sites block at least one crawler.
- Wrong if: 8 or fewer do.
- Inconclusive if: a site in the sample stops existing, or a crawler token is retired, so the check can’t be made as defined.
- Baseline: 9 of 10 on 2 October 2026.
How these get marked
On each check date I rerun the test and mark the call right, wrong or inconclusive on the predictions page, next to P1 and P2, which are already open and due on 2 January 2027. Wrong calls stay up.
Added 10 October 2026: these calls are P3, P4 and P5 on the scoreboard. They are marked hit, miss or void there, using the tests above unchanged. This post called those outcomes right, wrong and inconclusive.