What will change in AI search in 2027?

My five testable predictions: Search Console starts showing clicks from AI answers, pay per crawl stays a niche, checkout inside assistants stays invitation only for small shops, llms.txt still goes unread by assistants, and Google still sends at least ten times more clicks than every assistant combined on my own sites. Each one comes with the outcome that would prove it wrong by 31 December 2027.

Does Google Search Console show AI Mode and AI Overviews data?

Yes, since June 2026, as impressions. Search Console added generative AI performance reports covering AI Overviews and AI Mode, with breakdowns by page, country, device and date, but without clicks or query level detail. AI Mode rows gained a prompt filter in September 2026.

Does Google Analytics track traffic from ChatGPT?

It does when the visit arrives with a recognised referrer. In May 2026 Google Analytics added an AI Assistant channel to its default channel group, which groups sessions from assistants such as ChatGPT, Gemini and Claude. Visits from in app browsers or copied links still show up as direct traffic.

What Breaks in 2027: Five Predictions You Can Check

¶ MANIFESTOTHE RULE

01Most forecasts cannot lose

Every year produces a stack of pieces about where search is going, and almost none of them can be checked. Search will become more conversational. Brands will need to adapt. Trust will matter more. Each of those sentences will be true in 2027 whatever happens, which is exactly why they are worthless. A prediction without a way to be wrong is just a mood.

So this entry follows the rule this blog uses for every number. A claim has to carry the test that would break it. Each prediction below has a date, a specific outcome, and a line saying what would prove me wrong. On the first of January 2028 I will check all five and publish the score.

Two I nearly made, and lost already

When I drafted this list, the first two predictions were that Google would give Search Console a separate report for AI answers, and that Google Analytics would get a channel for traffic from assistants. Then I checked. Google shipped the Search Console report in June 2026, and the analytics channel in May. Both had happened before I could predict them. If you are not checking your predictions against this month, you are predicting last year.

// FIVE PREDICTIONS~~DUE 31 DEC 2027 · EACH ONE CAN FAIL~~

Search Console shows clicks from AI answersGoogle's AI report counts impressions only. I expect clicks to follow.

WRONG IFthe report still shows impressions only on 31 December 2027

Pay per crawl stays a nicheLicensing standards exist. I do not expect the largest AI companies to pay per crawl under one.

WRONG IFOpenAI, Google or Anthropic publicly agrees to pay sites per crawl or per answer under an open standard

Agent checkout stays invitation onlySelling inside an assistant will still need an application or a partner for a small shop.

WRONG IFany small merchant can switch on checkout inside a major assistant through self serve signup

llms.txt still goes unreadI expect no major assistant to say it reads the file when answering.

WRONG IFOpenAI, Google, Anthropic, Microsoft or Perplexity documents using llms.txt to answer questions

Google still out sends every assistantOn my own sites, search will still send at least ten clicks for every one from AI assistants combined.

WRONG IFmy analytics show assistant referrals above a tenth of organic search clicks in December 2027

I will check all five on the first of January 2028 and publish the score, whatever it is.

¶ MANIFESTOTHE REASONING

02Why these five

Measurement catches up, slowly

The first prediction follows from the pattern above. Bing put AI citations in its webmaster tools in February 2026, which I wrote about when I found the hidden Bing report. Google followed with impressions in June, and in September added a prompt filter for one of its AI surfaces. Clicks are the obvious next step, and also the most commercially awkward one, because they would show how little traffic the answers send back. That tension is why I gave it until the end of 2027 rather than predicting it for next month.

Two bets against my own wishes

The second and fourth predictions are bets against things I would like to happen. Licensing standards for paying sites per crawl already exist, and the largest AI companies have not committed to any of them. The file that tells machines what a site contains is cheap and sensible, and still, in my own logs, the llms.txt file is fetched by tools and directories far more than by any assistant answering a question. The useful forecasts are the ones that would disappoint you if they came true.

EXTRACTED — THE SENTENCE THIS ENTRY EXISTS FOR

Put a date on it and a way to fail, or keep it to yourself.

¶ MANIFESTOWHAT TO DO WITH THEM

03Plan for the ones you control

Commerce and the click

The third prediction is about buying inside assistants. One major assistant launched in chat checkout in September 2025 and scaled it back the following year. Google announced its own commerce protocol in January 2026 and was still running checkout as early access in September. My bet is that by the end of 2027 a small shop still cannot simply switch it on. The fifth is the one I can test myself: on my own sites, search will still send at least ten clicks for every click from all assistants combined.

There is a wrinkle in that last one that I should admit now rather than in January 2028. Clicks from Google’s own AI answers arrive as ordinary Google search clicks in most analytics, so if Google itself becomes the assistant, my test could pass while the thing it measures has changed underneath it. I am keeping the test exactly as written, because a prediction you can quietly redefine later is not a prediction, but the score will carry that caveat when I publish it.

What a reader should take from this

None of these predictions changes what to do on Monday. Being the clearest source on a narrow question works whether or not the clicks get reported, and so does being readable by every crawler. What they change is what to stop waiting for. If measurement is slow, keep your own logs. If paid licensing stays rare, do not plan revenue around it. If checkout stays closed, make your own checkout the easiest place to land. This is the same discipline as publishing claims with numbers, applied to the future. A forecast you can lose is worth more than ten you cannot, because only the first one can teach you anything. The score in January 2028 matters less than the habit of keeping one.

Tomorrow: sixty eight entries in, measured the same way as the first thirty, including what got worse.