What AI answers changed about SEO
The half of organic traffic that was never going to buy is disappearing. The half that was is getting harder to win and worth more when you do.
Most of the panic about AI search has been aimed at the wrong thing. Search was never going to disappear. What is disappearing is the least commercially useful half of your organic traffic, and too many content programmes were built entirely on that half.
If your organic strategy depended on ranking for questions a model can now answer without sending anyone anywhere, that strategy is finished. If it depended on being the source a buyer wants to check before spending money, it just got more defensible.
What is happening to traffic
Three effects, roughly in this order.
Impressions hold or rise while clicks fall, because the answer is being served in the results page. Top-of-funnel informational pages lose the most. Commercial and comparison pages lose the least. And the visits that still arrive convert at a noticeably better rate, because the people clicking now are the ones who need proof, pricing or a decision.
For a well-built site the net effect is less traffic and more revenue per visit. For a site built on programmatic volume it looks closer to a collapse.
Subjects rather than strings
The mechanical change underneath all of this is that search engines increasingly reason about things: companies, people, products, places. Not exact phrases.
That has practical consequences. It matters that your company is described consistently everywhere it appears. It matters that expert content carries a named author with verifiable credentials. It matters that your structured data is complete and accurate, because that is the most direct way to state what you are instead of hoping it gets inferred.
None of this is new advice exactly. What is new is that it moved from a minor ranking factor to the mechanism that decides whether you are quotable at all.
What still works, and works better than before
Original evidence. Proprietary data, benchmark studies, teardowns, methodology you can publish. Anything a model can generate by recombining existing pages has no reason to be cited. Anything only you could have written does.
Firsthand experience. Specific, dated, named accounts of having done the thing. The detail that would be embarrassing to invent is exactly the detail that reads as real.
Commercial pages. Comparison, alternative, pricing, integration and location pages sit closest to the transaction and are least likely to be answered without a click.
Technical fundamentals. Crawlability, rendering, speed and internal linking have not become less important. If anything they matter more, since a crawler that cannot cheaply parse your page will not include it in an answer.
What to stop doing
Publishing volume against informational queries with no commercial path. Rewriting the same competitor summaries everyone else has already published. Chasing a keyword because a tool gave it a low difficulty score, without asking whether the person searching it could ever buy from you.
And treating word count as a strategy. It was always weak. Now it is counterproductive, because thin derivative content drags down how the whole domain gets assessed.
Measuring it
Rankings and sessions are no longer enough on their own. Add three things.
Track share of voice across a defined set of commercially valuable topics against named competitors, which survives click-through volatility. Track citations in AI surfaces for your priority questions, sampled the same way each month. And track non-brand organic revenue, which is the only organic number that finally matters.
If those three are moving the right way, a fall in raw sessions is not a problem. It is just the shape of the channel now.