No — but the part of it that consisted of producing competent summaries of settled questions is finished. Search still sends enormous traffic to pages people want to open: comparisons, procedures with real detail, product and pricing pages, anything where the searcher has to evaluate rather than be told. What died is the middle of the market, where a page's only claim was that it ranked.

Are AI Overviews really reducing clicks?

For queries with one short factual answer, yes, and it shows up in Search Console as impressions holding while clicks fall. Treat the specific percentages with suspicion: the figures in circulation disagree widely and rarely state which query mix they measured, so a number from a definitional data set says nothing about a commercial one. Check your own Search Console by query type instead of adopting someone else's average.

Should I optimize for ChatGPT instead of Google?

It is mostly the same work, which is the good news. Answer-first structure, specific claims, named sources, clean markup and internal links serve both. The differences are matters of degree rather than a separate discipline — see generative engine optimization. Splitting the effort into two programmes tends to create duplicate pages competing with each other, which is a worse problem than either channel.

Do I need an llms.txt file?

It costs about an hour, so add it if you want to, but do not build a strategy on it. No major answer engine has committed to reading it as a retrieval or ranking input, and the pages getting cited today are cited because of what is on them. See llms.txt for what it is and what it currently is not.

Has AI search made SEO faster to pay off?

No, and arguably the opposite. A new site now has to be visibly worth citing before it earns anything, which is a higher bar than being merely relevant. Budget the same months you always would have, and judge progress on whether the pages you shipped are ones a model would have to attribute rather than absorb.

SEO & GEO

Is SEO still worth it now that AI answers the question?

By Jake Luo · Published 2026年7月30日

Yes, but the easy half of it is gone. AI answers absorb the queries that have one short factual answer, and the clicks those pages used to collect are not coming back. What still pays is the page a model has to quote rather than replace: specific, verifiable, and about a decision rather than a definition. The practical change is not doing less SEO — it is stopping the kind of page whose only asset was ranking for a question two sentences can settle.

What actually changed

Search used to end with a list of links, and an informational page's job was to be one of them. It now frequently ends with an answer: Google's AI Overviews render a synthesized response above the results, and a growing share of questions never reaches a search engine at all because it was typed into ChatGPT, Claude, Perplexity or Gemini instead. For a page whose entire value was ranking for such a question, that is a structural loss rather than a temporary dip. It is the zero-click search pattern arriving faster and across more query types than when featured snippets were the worst of it.

Be careful with the numbers attached to this story. Click-through declines are real and observable, but the percentages in circulation disagree with each other badly, and most of them never say which query set they measured. A study weighted toward definitional queries will report a collapse; one weighted toward commercial queries will report almost nothing; both can be honestly reported. What is safe to conclude is directional: the loss lands unevenly, concentrated on questions with a single settled answer, and barely touches queries where the searcher still has to compare, evaluate, buy, or read something in full.

Which pages lost, and which still earn

The useful way to audit your own site is not by traffic but page by page, asking whether an AI answer can substitute for it. If it can, that page was borrowing value from the search result and has none of its own.

The page that lost its value	The page that still earns
A definition and nothing else — the term, restated across three paragraphs.	A definition plus the decision it changes: when the term applies, and what it costs to get wrong.
A list assembled by reading the other results on page one.	A comparison you could only write because you have actually used the things in it.
A generic how-to for a task with one obvious method.	A procedure with the failure modes named, in the order they really bite.
A page whose single asset was its ranking.	A page a model can quote and a human still wants to open afterwards.

Notice the pattern down the right-hand column. Every one of them is harder to write and none can be produced by summarizing search results — which is precisely why a model cannot replace them, and why they also happen to be what a model is most likely to cite. Answer engine optimization and classic SEO have converged far more than the two labels suggest.

What to change in practice

None of this calls for a new discipline. It calls for four adjustments to how a page gets built:

Put the answer in the opening. Not after the context, not after the history of the term — the first two to four sentences, phrased so they survive being lifted out of the page. That is the part an AI answer quotes and the part a skimming human reads, and both want the same thing.
Get narrower. A page about a broad topic competes with everything and gets summarized; a page about one specific situation is often the only thing that addresses it. Specificity is the whole argument for the long-tail keyword, and prompts run longer than queries, so it matters more now than it did.
Put something on the page that cannot be summarized from elsewhere. A number you measured, a decision you made and why, a failure you had. This is the only durable defence: consensus content is what a model synthesizes, first-hand content is what it has to attribute.
Stop treating rank as the only outcome. Being described accurately inside an AI answer is worth something even when it sends no click, and it is invisible in a rank tracker. Watch branded search and direct traffic alongside positions — see how do I measure if my marketing is working.

First-party note from building AgentCeres — the AI Growth Officer at agentceres.com: we run this site's answer pages as a working system, and two things are worth passing on. The first is structural. The direct answer to each question is stored as its own field, separate from the body, so it cannot drift into the middle of a page where nothing will lift it; treating the lead as a distinct artefact rather than an introduction is most of what answer-first means in practice. The second is a limitation nobody selling GEO services mentions: no report anywhere tells you an AI engine cited you. Search Console shows queries, analytics shows sessions, and neither has a row for quoted in an answer with no click. The loop does not close. You can verify that a page deserves citing — specific, sourced, quotable — and you cannot verify that it was cited, which means the only defensible strategy is to make pages worth citing and stop waiting for a dashboard to confirm it.

When SEO is the wrong bet anyway

Separately from AI, SEO is still the slowest channel a founder can pick, and that has not changed. It compounds over months, it is indifferent to your runway, and a brand-new domain spends a long stretch earning the right to rank at all — see how long does SEO take to work. If you need users this quarter, the answer is outreach, a launch, a community, or paid, and the AI question does not enter into it.

The reasonable version is to run SEO as the compounding channel behind a faster one rather than instead of one, and to size it honestly: a few pages genuinely worth citing beat fifty that summarize what is already indexed. If you are choosing between channels rather than sequencing them, how do I know which marketing channel to double down on is the more useful question — and if the answer is still SEO, how do I do SEO for a brand new website covers the starting point.

FAQ

Is SEO dead?: No — but the part of it that consisted of producing competent summaries of settled questions is finished. Search still sends enormous traffic to pages people want to open: comparisons, procedures with real detail, product and pricing pages, anything where the searcher has to evaluate rather than be told. What died is the middle of the market, where a page's only claim was that it ranked.
Are AI Overviews really reducing clicks?: For queries with one short factual answer, yes, and it shows up in Search Console as impressions holding while clicks fall. Treat the specific percentages with suspicion: the figures in circulation disagree widely and rarely state which query mix they measured, so a number from a definitional data set says nothing about a commercial one. Check your own Search Console by query type instead of adopting someone else's average.
Should I optimize for ChatGPT instead of Google?: It is mostly the same work, which is the good news. Answer-first structure, specific claims, named sources, clean markup and internal links serve both. The differences are matters of degree rather than a separate discipline — see generative engine optimization. Splitting the effort into two programmes tends to create duplicate pages competing with each other, which is a worse problem than either channel.
Do I need an llms.txt file?: It costs about an hour, so add it if you want to, but do not build a strategy on it. No major answer engine has committed to reading it as a retrieval or ranking input, and the pages getting cited today are cited because of what is on them. See llms.txt for what it is and what it currently is not.
Has AI search made SEO faster to pay off?: No, and arguably the opposite. A new site now has to be visibly worth citing before it earns anything, which is a higher bar than being merely relevant. Budget the same months you always would have, and judge progress on whether the pages you shipped are ones a model would have to attribute rather than absorb.

Want this done for you?

AgentCeres is a managed AI marketing team — specialists draft the work, you approve what ships. 14-day free trial, from $39/month.

Start free trial More answers