GEO vs SEO in 2026: what can actually be measured, and what cannot

Generative Engine Optimization has a credibility problem. The underlying shift is real: people ask AI assistants for recommendations and read a synthesised answer with a handful of citations. But a lot of what is currently sold as “GEO” cannot be measured at all, and some of it is simply SEO with a new invoice.

This piece separates the two.

What actually changed

Classic SEO optimises for a ranked list of links. You can see your position, your impressions, and your clicks in Search Console. The feedback loop is tight.

AI answers work differently. The engine reads sources, synthesises an answer, and cites a few of them. There is no “position 3”. Often there is no click at all, because the user got what they needed inside the answer.

That single difference — no guaranteed click — is why measurement is harder, and why anyone promising you precise GEO rankings should be treated with suspicion.

What genuinely can be measured

1. Citation presence across a fixed prompt set. Write 30 to 50 prompts your buyers would realistically ask (“best [service] provider in [city]”, “[competitor] alternatives”, “how much does [service] cost”). Run them monthly across the major assistants and record whether you are cited. It is manual and imperfect, but it is a real, repeatable measurement, and the trend line is meaningful.

2. Crawler access. Whether GPTBot, ClaudeBot, PerplexityBot, Google-Extended and others can reach your content is a binary fact you control. You can verify it in robots.txt, headers, and server logs.

3. Referral traffic from AI assistants. Perplexity, ChatGPT and others do send referral traffic, and it appears in analytics with identifiable sources. It is usually small compared to search, but it is real, attributable, and worth isolating in its own report.

4. Structured data validity. Whether your Organization, Service, Offer, FAQ and Breadcrumb markup is present and valid is testable and binary.

5. Brand mention footprint. Whether you appear in the third-party sources AI systems tend to lean on — directories, comparison posts, roundups, reviews — can be counted and grown deliberately.

6. Classic SEO metrics, still. Impressions, clicks, and average position have not stopped mattering. Google’s AI Overviews still draw heavily on the same web graph.

What cannot honestly be measured

Be sceptical of anyone claiming these:

• “Your GEO ranking is #2.” There is no stable ranking to occupy. Citation varies by phrasing, user, session, and model version.
• “We increased your AI visibility 300%.” Against what baseline, on which prompts, on which model, on which date? Without a published prompt set and method, this is decoration.
• “We got you into the training data.” Nobody outside the labs controls this, and training cut-offs are not a service anyone can sell you.
• Precise attribution of a sale to an AI answer. When there is no click, there is no referrer. A customer who says “ChatGPT recommended you” is genuinely useful evidence, which is why you should simply ask on your enquiry form.

The overlap nobody mentions

Much of what makes content citable is what made it good SEO in the first place: it is crawlable, well structured, factually specific, clearly attributed, and genuinely useful. The GEO-specific additions are real but narrower than the marketing suggests:

• Writing passages that are directly quotable — short, specific, self-contained answers rather than 400 words of throat-clearing.
• Making facts explicit and attributable, including prices, timelines and specifics an assistant can safely repeat.
• Ensuring crawler access for AI user agents you want to be read by.
• Building third-party corroboration, because assistants weight sources that other sources agree with.
• Publishing an llms.txt summary. Worth doing, cheap, and honestly: not a ranking lever. Anyone selling it as one is overselling.

How to judge whether GEO work is paying off

Set this up before you start, or you will be arguing about vibes in three months:

1. Publish the prompt set. Fixed, written down, unchanged between measurements.
2. Baseline it. Record citation presence for every prompt before any work begins.
3. Re-run monthly, same prompts, same method, and keep the raw results.
4. Track AI referral traffic separately in analytics.
5. Ask new enquirers how they found you, and include an AI assistant option. This is the single most underrated GEO measurement, and it is free.
6. Watch the classic metrics too, because most of the technical work benefits both.

If citation presence on your prompt set rises over three to six months, AI referral traffic appears and grows, and people start telling you an assistant mentioned you — that is a real result.

The honest bottom line

GEO is worth doing, especially early, because citation patterns are sticky and the field is not yet crowded. But it should be sold and bought as structural work with a measurable trend line, not as a ranking product with precise numbers.

If a provider cannot show you their prompt set, their baseline, and their method, you are not buying measurement. You are buying a story.

See what our GEO engagement actually includes and how we report it: Generative Engine Optimization.