Why SEO changed with AI Overviews

The first thing most people notice about AI Overviews is the traffic. An informational page that used to sit at position two and pull a steady stream of clicks now sits behind a generated answer, and the clicks thin out. That part is real, and pretending otherwise doesn’t help anyone.

But the conclusion people draw from it — that SEO is finished — doesn’t follow. What actually changed is narrower and more useful to understand.

The queries that lost clicks were never the valuable ones

Look at which pages lost. It is almost always the definitional, top-of-funnel stuff: “what is X”, “how does Y work”, “X vs Y” at the most basic level. Those queries were always poor converters. They filled dashboards and they made traffic charts look healthy, but the person asking “what is technical SEO” is rarely the person about to hire someone to do it.

Queries with intent behind them — commercial comparisons, pricing, local services, anything where the searcher needs to pick a provider rather than learn a definition — still send clicks, because a generated paragraph doesn’t finish the job. Somebody deciding who to hire has to visit sites and form a judgement.

So the first move is unglamorous: re-audit what your traffic was actually worth. If a third of your sessions came from queries that never produced an enquiry, losing them is a reporting problem, not a business one.

Being the source is the new position one

AI Overviews cite. Those citations are a real acquisition channel, and they behave differently from blue links — being cited depends less on classic authority signals and more on whether your page states a clear, extractable, verifiable claim near the question it answers.

In practice that means:

• Answer the question in the first paragraph under the heading. Not three paragraphs of context first. The extractable answer should be adjacent to the heading that matches the query.
• Be specific enough to be quotable. “Improves performance” doesn’t get cited. “Reduced LCP from 4.1s to 1.6s by self-hosting the font” does.
• Keep one claim per passage. Dense paragraphs that bundle four ideas are hard to extract cleanly.
• Give it something to attribute. Original data, a named methodology, a first-hand result. Synthesised summaries of what everyone else already wrote have no reason to be picked as the source.

Structured data is doing more work, not less

Schema was always a hint rather than a ranking factor, but it has become a much more reliable way to state unambiguously what a page is. Article, FAQPage, ProfessionalService, Product with real offer data — these make a page cheaper to interpret correctly, and cheap-to-interpret pages get used.

The same goes for the boring hygiene: clean heading hierarchy, real HTML semantics, tables that are actually tables. Every shortcut that made a page harder to parse now costs more than it used to.

What I’d actually change

If you are running a site today, the honest priority order looks like this:

1. Segment your reporting by intent before you panic about a traffic decline. Know what you lost.
2. Move budget from definitional content toward commercial and comparison content, where clicks still convert.
3. Restructure existing pages so the answer sits directly under the matching heading.
4. Add the structured data you have been meaning to add.
5. Publish something only you could publish — your data, your results, your method.

None of that is a new discipline. It is the same work, re-prioritised, with the low-value half of the old playbook finally exposed as low-value.

The sites getting hurt worst are the ones whose entire strategy was ranking for questions they had no particular authority to answer. That was always fragile. AI Overviews just sent the invoice.