Every week someone asks me to skip to the interesting part. They have read that Google’s organic click-through rate is falling, they have seen their brand fail to appear in ChatGPT, and they want a GEO engagement — no audit, no technical remediation, straight to getting cited by the models.
I turn those down, and this is the argument I make first.
The uncomfortable mechanical fact
Generative engines do not have a private, superior index of the web. Whatever their architecture, in production they are overwhelmingly reading pages that were discovered, fetched, parsed and ranked by conventional retrieval — their own crawlers, a search partner’s index, or a hybrid. The retrieval step is still retrieval.
Which means the failure modes stack in a specific order:
- If your page cannot be crawled, nothing downstream happens.
- If it renders its content only in client-side JavaScript, many retrieval pipelines see an empty shell.
- If it is crawled but ranks nowhere for the topic, it is rarely in the candidate set the model gets to read.
- If it is in the candidate set but structurally unreadable, it does not get quoted.
- Only at that fifth layer do the interesting GEO questions begin.
People want to work on layer five. Almost every underperforming site I audit is failing at layers one through three.
What “SEO first” concretely means in my process
It is not a stalling tactic and it is not a bigger invoice. It is a fixed list, and it is usually short:
Crawl and index integrity. Can Googlebot and the AI crawlers reach the pages that matter? Is the robots.txt accidentally excluding them — including GPTBot, ClaudeBot, PerplexityBot and Google-Extended, which I now check explicitly, because I have found more than one client blocking every AI crawler while paying someone to improve their AI visibility.
Rendering. Does the content exist in the served HTML, or only after hydration? This single issue silently disqualifies more pages from AI citation than any content problem.
Cannibalization. If four of your pages half-answer the same question, retrieval has four mediocre candidates instead of one strong one. Classical search handles this badly. Generative retrieval handles it worse, because there is no second result — there is one answer, assembled, and your four pages compete to be a fragment of it.
Entity clarity. Does the site state plainly who it is, what it does, and where? Organization and Person schema, a real About page, consistent naming. Models are entity-resolution machines, and ambiguity about who you are is fatal at the citation step even when the content is good.
Baseline quality and trust signals. Author attribution, dates, sourcing, and for anything YMYL, credentials that a reasonable evaluator could verify.
That list is unglamorous. It is also where the traffic is.
Then, and only then, GEO
Once the foundation holds, the optimization genuinely changes, and this is the part worth being excited about.
Answer-first structure. Lead each section with a direct, self-contained, extractable answer of one to three sentences, then expand. Models lift passages, not pages. A passage that only makes sense after reading the two paragraphs above it will not be lifted.
Semantic completeness over keyword density. The question is no longer whether you used the phrase. It is whether you covered the entities and sub-questions that co-occur with the topic across the corpus. Gaps in coverage are why a well-ranking page still loses the citation to a thinner competitor that happened to address the specific sub-question asked.
Question-shaped headings. Real user phrasing, not marketing phrasing. “How much does X cost in Egypt?” retrieves; “Pricing” does not.
Explicit, quotable claims with sources. Models preferentially cite text that carries its own evidence. A sentence containing a number, a date and an attribution is worth more than a paragraph of confident assertion.
Consistent facts across every surface. Your site, LinkedIn, Crunchbase, directory listings, Wikipedia if you are there. When sources disagree about your founding year or your service list, the model’s confidence drops and it reaches for a competitor it can state cleanly.
What I tell clients about measurement
GEO measurement is genuinely immature, and I would rather say so than sell certainty. Rank tracking does not apply — there is no position three. What I actually track:
- Prompt-set monitoring: a fixed list of 30–60 real buying-intent prompts, run on a schedule across ChatGPT, Gemini, Perplexity and Claude, logging whether the brand appears and whether it is cited or merely mentioned.
- Referral traffic segmented by AI source in analytics, which is undercounted but directionally useful.
- AI crawler hits in server logs — the leading indicator that anything is working at all.
- Branded search volume, which tends to move before anything else does.
None of these is a clean KPI. Anyone showing you a single confident “GEO score” has invented it.
The honest summary
GEO is real, it matters, and the shift in click behaviour is not hype. But it is a layer on top of search infrastructure, not a replacement for it, and the agencies selling it as a standalone product are mostly selling the same technical SEO audit with a new cover page — or worse, skipping the audit.
Fix the foundation. Then optimize for the answer. In that order, every time.
Amgad Salem leads SEO at Eye Advertising and teaches the AI-Assisted SEO & GEO programme at Learn N’ Digital. See his SEO and GEO case studies, or follow his work on LinkedIn.
Leave a Reply