Search visibility in the AI era: from SEO to GEO

The answer to a search query is increasingly a paragraph rather than a list of links. Google AI Overviews, ChatGPT's search mode, Perplexity and Claude all do the same thing: read sources, synthesise them, and hand the user one answer. This does not invalidate the decade-old reflex of "rank first for the keyword" — it just makes it insufficient on its own.

This article explains GEO — Generative Engine Optimization — where it diverges from classic SEO, and what a corporate site should actually do about it.

What exactly separates SEO from GEO

Both systems aim at the same thing: getting the right content to the right person. But the surface where the answer appears is different, so the optimisation target shifts.

Classic search: ten blue links

In classic search the competition is about ranking. Reach position one and you get the click. The unit of optimisation is the page: title tag, description, internal linking and authority signals all operate at page level.

Generative search: one answer, a handful of sources

In generative search the competition is about being cited. While composing an answer, a model pulls passages from three to eight sources. The unit of optimisation is no longer the page but the passage: a block of text the model can lift and place into its answer without it losing meaning.

The practical consequence: a long, flowing article where every paragraph depends on the previous one reads beautifully to a human and is close to useless to a model.

How click behaviour is shifting

The chart below is not a measured dataset — it is a directional scenario model: a simplified picture of the shift we typically observe on informational queries. You need to derive your own site's real distribution from Search Console and referrer data.

The point is not the percentages but the widening gap. A site that holds exactly the same rankings can lose traffic while doing nothing wrong. What compensates for that loss is the visibility gained in exchange: being named inside the answer.

How to structure content for GEO

Five rules, in the order you should apply them.

1. Answer in the first paragraph

The first two sentences under each heading must answer the question that heading poses, on their own. No lead-in, no context-setting, no warm-up. The model reads the first passage and takes it if it is sufficient.

2. Write headings as questions

People ask language models in full sentences. The closer your headings are to those sentences, the stronger the match. Use "How much does a corporate website cost" instead of "Pricing", "How long does delivery take" instead of "Process".

3. Add quotable data blocks

Tables, numbered lists and concrete numbers are the easiest units for a model to carry. Turning a claim into a three-row table instead of defending it in prose measurably raises the odds of being cited.

4. Implement schema.org fully

Structured data works both for classic rich results and for entity extraction by models. The minimum set on a corporate site:

Every schema above is live on this site; we detail the approach on our SEO & GEO service page.

5. Signal freshness and sources

When models see conflicting information, they look at dates. Publish and update dates should be visible, numbers should state the period they refer to, and external sources should be named. Undated content is the first thing dropped on questions where recency matters.

The technical base: making the page readable

However good the content is, it means nothing if the page cannot be read. Three checks:

Server rendering. Content that only materialises after JavaScript runs in the browser is a blank page to many AI crawlers. In a Next.js site the rule is simple: every page that does not touch a database should be statically generated. We covered the performance side of that in how to build a fast website.

robots.txt. Many sites block GPTBot, ClaudeBot or PerplexityBot without realising it. Blocking them can be a legitimate choice — but it should be a deliberate one. If you want visibility, these agents need access.

Speed. Crawl budget is not a Google-only concept. Slow pages get crawled less, and pages crawled less get cited less.

# robots.txt — for a corporate site that wants visibility
User-agent: GPTBot
Allow: /
 
User-agent: ClaudeBot
Allow: /
 
User-agent: PerplexityBot
Allow: /
 
Sitemap: https://example.com/sitemap.xml

How to measure GEO

Unlike classic SEO there is no central dashboard. Measurement runs on three tracks.

Manual query testing is the most practical method: ask your target questions to four or five different models at regular intervals and record which sources appear. Even twenty queries a month produce a meaningful trend within a quarter.

Referrer analysis means defining AI-sourced referrals as a separate segment in your analytics. That traffic usually shows above-average session duration and conversion rates, because the visitor arrives already informed.

A 90-day implementation plan

The sequence mirrors how a corporate web project actually runs: foundation first, then content, measurement last. Breaking that order — producing content before the technical base is ready — is the most common waste of time we see.

Five common mistakes

1. Chasing keyword density. Models look at semantic coherence, not term counts. Repeating a word ten times helps nobody.
2. Publishing generic AI-written content as-is. Models do not cite the average text they themselves produce; what stands out is original data, examples and experience.
3. Piling everything into one long page. Passage-level citation rewards clearly segmented structure.
4. Hiding concrete information like price, timeline and scope. That is exactly what models look for; "contact us for details" never gets quoted.
5. Staying monolingual. On a multilingual site each language is a separate visibility surface; with hreflang set up properly the cost is low and the return is high.

Seeing it in practice

On Ambalaj Cini we ran development and ad management together, and the company grew 30%. For what is possible on the technical side, Nexos Investment is the better example: multilingual structure, schema, server rendering and a performance budget in one project. The rest of our work is on the projects page.

Conclusion

GEO is not a fad that invalidates SEO; it is a technical response to search results changing shape. Most of the work is what good SEO already demanded: fast, server-rendered pages, a clear heading hierarchy, correct schema, honest and concrete content. The only thing that changed is that we now do it for a second audience too — one that reads passage by passage rather than sentence by sentence.

If you are not sure which of those two readers your site is ready for, write to us and we will review the current state together.