You type your brand name into ChatGPT or DeepSeek. The answer names three competitors and not you. Then you open Search Console and your money keywords are still sitting on page one. That gap is the evidence behind every "SEO is dead" headline you have read this year.
The more accurate conclusion: SEO is not dead, it has been demoted to infrastructure. It gets you crawled, understood and trusted. Being selected and quoted inside an AI answer is a separate capability, usually called GEO. This article separates the two layers: what still works as before, what you have to add, how to sequence the work, and how to measure it.
The short answer: SEO gets you in, GEO gets you quoted
SEO makes your pages crawlable, indexable and trustworthy. GEO makes an AI willing to use those pages as source material. One decides entry, the other decides whether you appear.
The two are not substitutes. Without crawlable pages, indexable body text and verifiable authority signals, a model cannot even read your content, let alone cite it. But a site with solid technical foundations and no liftable answer sentences, no consistent entity description, will rank in search results and still never make it into an answer. Once you see the stack this way, "is SEO still worth it" stops being a binary question.
What actually changed, and what did not
The change is not in crawling or indexing. It is in the shape of the results page: ten blue links became a paragraph with two or three citations, and clicks dropped accordingly.
That reshapes how content earns its keep. An article used to pay off through ranking plus clicks. Now there is an extra filter: a model first decides whether a passage answers the question directly, then decides whether to quote it. The writing goal shifts from "get the user to click" to "get the answer to quote you, so the user follows the citation back."
Google's official guidance on AI features states that sites do not need special optimization for AI search and do not need to create new machine-readable files, because existing SEO fundamentals still apply. That sentence is often read as "GEO is a myth." What it actually says is: the foundation is unchanged, the layer above it has to be rebuilt. The foundation decides whether you are eligible; the upper layer decides whether you get named.
The part of SEO that still carries the load
Crawling, indexing, page experience and authority signals remain the precondition for every AI answer. None of that work has expired.
- Crawl permission: robots.txt still works as specified in RFC 9309. AI crawlers such as GPTBot and PerplexityBot honor it, which means a blanket block is the most common self-inflicted wound in this space.
- Indexable body text: if your content is client-side rendered and your key paragraphs only exist inside images, no engine can read them, so none can cite them.
- Structured data: schema.org markup tells a machine what the entity is, who wrote it and when it changed. That is the entry ticket to being understood.
- Authority and consistency: one entity, one name, one positioning sentence, repeated everywhere. The more independent sites reference it, the more likely a model treats it as a credible source.
This is classic SEO, and it happens to be the foundation GEO stands on. Abandoning it to chase "AI optimization tricks" is building on sand: nothing breaks immediately, then one crawl or ranking change takes the whole thing down. For the underlying mechanics, the GEO primer covers how retrieval and citation actually work.
What GEO adds: content that can be lifted
GEO has one core move: write expertise as answer sentences an AI can lift intact, and make those sentences verifiable across the web.
In practice, four techniques do most of the work:
- Conclusion first: the opening paragraph of every section is a complete 40–80 word conclusion. No warm-up, no suspense, no background essay.
- Definition sentences: give core concepts the standard "X is..." form. These are among the most frequently quoted sentence types in AI answers.
- Structured comparison: put "A versus B" into a table. Machines parse tables far more reliably than flowing prose.
- Traceability: every number carries a source, and if there is no source, there is no number. This also protects brand credibility.
One more rule gets overlooked: entity consistency. When a model decides whether two pages describe the same company, it looks for a repeated name, positioning line, contact detail and product naming across the web. Publishing as "AiVisible" on your site and a translated variant elsewhere splits you into two entities, and neither accumulates full trust.
Who does what: a side-by-side view
Put both layers in one table and the priority order becomes obvious: SEO work cannot be skipped, GEO work cannot be half done.
| Work item | SEO layer | GEO layer |
|---|---|---|
| Crawling | Allow bots, submit sitemaps | Confirm AI crawlers are not blocked, keep logs |
| Content shape | Keyword coverage, internal links | Conclusion sentences, definitions, comparison tables |
| Structure | Heading hierarchy, semantic HTML | Entity markup, FAQ structure |
| Authority | Backlinks, branded search | Cross-platform entity consistency, third-party mentions |
| Measurement | Rankings, clicks, CTR | Mention rate, citation position, landing page mix |
Look at the last row. The two layers are measured differently. Judging GEO by rankings is like judging an email campaign by download count: the metric and the outcome do not line up, and the team argues about it forever. It is also why many companies finish an "AI optimization" project without being able to say whether it worked.
Why SEO-only sites disappear from answers
SEO-only sites rarely lack rankings. They lack liftable answer sentences, which is the most common failure point.
Three patterns worth checking against your own site:
- The answer is buried: one 400-word paragraph carries three ideas, so a model cannot extract a clean sentence and skips the whole block.
- Product talk, no problem talk: the page is a feature list with no "use this when..." framing, while buyers ask exactly that.
- No third-party corroboration: the site only quotes itself, with no independent mention a model can cross-check.
None of these require a full rewrite. Three to five core pages usually move the needle. The catch is that you have to know how AI currently describes you before you edit, and guessing is not a measurement.
A six-step plan you can run in two weeks
Running SEO and GEO as one program takes six steps, and the first pass fits into two weeks.
- Baseline first: ask 20 real customer questions across your target platforms. Record whether you appear, at what position, and who gets cited instead. This is the anchor for every later comparison.
- Fix the foundation: allow AI crawlers in robots.txt, serve body text server-side, add Organization and Article structured data.
- Align topics to questions: move from keywords to questions. One H2 per real question, titled close to how people actually ask it.
- Rewrite key pages: add a conclusion sentence, a definition and a comparison table per section. Delete warm-ups and adjective stacks.
- Build external mentions: consistent entity descriptions in trade media, vertical directories and partner pages. These are the corroborating signals models use for authority.
- Re-test monthly: rerun the same question set, compare mention rate and position, then decide which pages to rewrite next.
Do not invert the order. Foundation first, content second, is the lowest-cost order to work in. Content first means writing into a directory nobody can read.
How to measure: from rankings to citations
Change the measurement layer: keep watching rankings, but make the primary metric the share of target questions where you are mentioned, and where you rank in the answer.
Track four numbers together: brand mention rate on your target question set, average citation position, the distribution of landing pages being cited, and organic traffic to those pages. The first three answer "does AI recognize us," the last answers "does recognition turn into pipeline." Only together do they tell you whether to invest in more content or fix technical issues first.
This used to be manual screenshot work. Tools now track it continuously. AiVisible, for example, turns those four numbers into a tracked set: which platforms, which questions, which position, and which page was cited, alongside an AI visibility score for trend lines. The score is not the goal; it makes progress discussable. Platform differences deserve their own review too, since citation behavior on DeepSeek does not match Western engines.
FAQ
Will GEO replace SEO?
No. GEO depends on the crawling, indexing and authority that SEO provides, so the two are layers of one stack. Running GEO without fixing the technical foundation leaves your content unreadable to models. Treat them as one program with one schedule rather than a choice.
Do I need a special file like llms.txt for AI search?
It is not required. Google's guidance states that sites need no AI-specific files and that existing SEO fundamentals still apply. The llms.txt proposal is still a community convention; it can supplement your setup but cannot replace robots.txt and indexable body text.
How long until AI answers change?
Measure in months rather than days, since timing depends on crawl frequency and how often you publish. Lock a set of 20 questions as a baseline and rerun the same set monthly instead of drawing conclusions from a single check. Citation changes usually show up one or two crawl cycles after a rewrite.
Should Chinese and Western AI platforms be handled separately?
Separate the data, share the structure. DeepSeek, Doubao and Tongyi differ from ChatGPT and Perplexity in what they crawl and quote, so the same article needs separate verification. Start with the one or two platforms where your buyers actually are.
With a limited budget, what comes first?
Fix the foundation before writing. Blocked AI crawlers and unindexable body text are cheap to repair and affect everything downstream, so they come first. Next, rewrite three to five core pages with conclusion sentences and comparison tables; those two steps often change how AI describes you.