The answer moved. The index didn't.
Here's the mechanic that most of the panic skips over. When Google generates an AI Overview, it isn't consulting a separate, secret pool of AI-approved content. It retrieves from the same web index that produces the ten blue links, applies broadly the same quality and relevance signals, then writes a summary over the top with citations attached.
That has one consequence worth more than every other take on the subject: the pages that get cited are, overwhelmingly, pages that were already ranking for the query or something adjacent to it. Being crawlable, being indexed, being genuinely useful on the query and having enough external corroboration to be trusted — the entry requirements didn't change. The layout on top of them did.
So the honest framing isn't "SEO is dead, learn GEO". It's that ranking became necessary but less sufficient. You still have to do the ranking part. You just can't stop there. What Google AI Overviews are covers the mechanics in more detail.
The click compression, with the actual numbers
Two pieces of published research are worth knowing, and both deserve their caveats stated out loud.
Pew Research Center, publishing in July 2025 from browsing data collected in March 2025, found that when an AI summary appeared, users clicked a traditional search result on 8% of visits — against 15% of visits where no AI summary appeared. Roughly half the click rate. They also found that around 1% of visits produced a click on a source link inside the summary itself, which tells you the citation is a credibility signal far more than a traffic source.
Ahrefs published its own analysis in 2025 reporting a substantial average click-through-rate drop for the top-ranking page on queries where an AI Overview appears — in the region of a third.
Both studies are directional rather than universal. Pew's panel is US adults, the sample is modest, and search behaviour in India differs on device mix, language and query length. Ahrefs' data skews to the keyword sets its customers track, which are not a random sample of the web. Treat them as evidence that the effect is real and material, not as a coefficient you can apply to your own traffic.
Google has said publicly that clicks arriving via AI Overviews are higher quality. It has not published data supporting that. Weigh it accordingly.
What got hollowed out, and what barely moved
The loss isn't spread evenly, and understanding the pattern is what turns this from a scary headline into a plan.
- If your content strategy was built on definitional top-of-funnel volume, the ground genuinely shifted under you.
- If it was built on the queries people type with a credit card nearby, you lost far less than the headlines suggest.
- The strategic move is not "write more". It's to shift the centre of gravity of what you publish down the funnel.
| Query type | Example | What happened |
|---|---|---|
| Definitional | "what is a canonical tag" | Hardest hit. The answer fits in three sentences and the AI writes those three sentences. |
| Simple factual | "gst rate on services" | Largely absorbed. There was never much of a click here anyway. |
| Comparison and evaluation | "agency vs freelancer for seo" | Partly summarised, but people still click through to check the reasoning. Trust matters more than speed here. |
| Commercial and transactional | "seo agency pricing india" | Least affected. Somebody about to spend money wants to see the actual page, the actual number and the actual company. |
| Branded | "[your company] pricing" | Barely touched. Nobody asks an AI to summarise a brand they've already decided to evaluate. |
Branded demand became the asymmetry that decides everything
Branded search is the one query class no summary can take from you. When somebody types your company name, they've already chosen to evaluate you, and no AI answer talks them out of the click.
This creates an asymmetry that's been quietly widening. A brand people already know gets cited more often (familiarity and corroboration across the web are both signals), gets clicked more often when cited, and converts better when clicked. A brand nobody knows gets summarised and skipped.
Which means the SEO work that pays best now is the work that also builds recognition: original data nobody else has, named authors with real credentials, opinions that can be quoted, and being present in the places your category gets discussed. That's a meaningful shift from the old model of covering every keyword permutation and waiting.
The practical test: track your monthly branded search volume as a demand indicator. If unbranded clicks fall but branded searches climb, the channel is working and your dashboard is lying to you.
The cost curve, which is the argument that never depended on AI at all
Set the AI question aside for a moment. There's a structural difference between organic and paid that's been true for twenty years and gets more pronounced as ad auctions get more crowded.
| Dimension | Paid search and social | SEO |
|---|---|---|
| Cost of the next click | Set by auction. Rises as competitors bid and as your audience saturates. | Falls as the page ages and accumulates authority. The marginal click is close to free. |
| When you stop paying | Traffic stops the same day. There is no residual. | Decays over months to years. Pages you wrote in year one still land leads in year three. |
| What you're left holding | Ad account data and learnings. Useful, not an asset. | Pages, links, rankings and a brand people recognise. These sit on your balance sheet in every sense but the accounting one. |
| Speed to first result | Days. This is the real advantage and it's a big one. | Three to six months for honest movement, six to twelve to compound. |
| Failure mode | You overspend and find out fast. | You underinvest and find out slowly. This one is more dangerous. |
What has to change in how you report this
Most SEO reporting was designed for a world where an impression reliably produced a click at a known rate. That relationship broke, and reports that ignore it will show failure where there is none — or hide failure where there is.
Five adjustments, in rough order of urgency:
- Stop leading with sessions. Lead with qualified leads from organic, tracked in your CRM. Everything else is an intermediate variable.
- Read impressions and clicks as two separate stories. Rising impressions with flat clicks used to mean a CTR problem. Now it can also mean your page is being read inside an answer box. Both are possible; find out which before you rewrite your title tags.
- Segment CTR by query type, not sitewide. As things stand, Search Console folds AI Overview appearances into the same impression and click counts, so you cannot cleanly split them in the interface. Segmenting by definitional versus commercial queries is the closest honest proxy.
- Track branded search volume monthly. It's the cleanest available read on whether the work is building demand rather than just harvesting it.
- Sample your citations manually. Take your twenty most valuable queries, run them through the assistants your buyers use, and log whether you appear. The tooling for this is immature and mostly overpromises; a spreadsheet and forty minutes a month beats a dashboard that guesses. Generative engine optimisation covers what's real here and what isn't.