Most advice about Generative Engine Optimization is about writing. Write shorter paragraphs. Write definitions. Write question-shaped headings. That advice is not wrong, but it puts the work in the wrong order — and it is why so many teams spend a month rewriting pages and still never see their name inside an AI answer.
I want to argue the opposite starting point. GEO is an extraction problem before it is a writing problem. If a model cannot lift a clean, self-contained sentence out of your page, no amount of prose polish helps. So the first job is to find out what your page currently gives a model to work with.
What an answer engine is actually doing
An AI assistant answering "what is the best CRM for a small business in Vietnam" is not ranking ten blue links. It is assembling a paragraph out of fragments it has already collected, and it has to decide which fragments are worth trusting.
Five things push a fragment up that decision:
- Topical fit and recency. A page about the exact question, updated recently, beats a broader page from two years ago.
- Parseability. Headings, lists, tables and short paragraphs survive extraction. Walls of marketing copy do not.
- Quotability. A sentence that stands alone — subject, claim, no dangling pronouns — can be lifted. A sentence that begins "As we mentioned above" cannot.
- Authority markers. Named authors, cited sources, credentials, inbound links.
- Entity clarity. The engine can tell who you are, where you operate and what you sell.
That list is a checklist you can audit against. It is not a style guide.
The audit I run before touching a draft
Four passes, in this order. Each one produces a concrete defect list rather than an opinion.
Pass one: can a machine see the structure?
Open the page with JavaScript disabled, or read the raw HTML. If your comparison lives in styled divs instead of a real <table> with a header row and units, the model gets a soup of numbers with no labels. If your key facts are baked into a screenshot, they effectively do not exist for extraction purposes.
The fix is boring and fast: one H1, a logical H2/H3 hierarchy, real tables, key facts in text.
Pass two: is any single sentence quotable?
Take your five most valuable pages and read the opening paragraph of each. If it opens with a brand statement or a promise ("we help businesses transform"), there is nothing to quote. Rewrite the opening as a direct answer in 40–60 words — the answer first, the positioning after.
Then read the rest aloud. Paragraphs longer than three sentences are extraction hazards: the model has to decide which part of the paragraph is the claim. Split them.
Pass three: does the schema match the page?
This is where most sites quietly fail. A page can carry FAQPage JSON-LD whose questions do not match the visible FAQ word for word — usually because the schema was pasted from a template and the visible text was edited later. Google tolerates it poorly and an answer engine has no reason to trust either version.
The test is mechanical: diff the visible question text against the schema text. Zero differences, or fix it. The same applies to Article, BreadcrumbList and Organization blocks — they should describe the page that exists, not the page you planned.
Pass four: can you measure whether any of this worked?
Here is the uncomfortable part. You can do passes one through three perfectly and have no idea whether an AI engine ever cited you, because classic analytics were not built for this.
Three signals are worth watching, and none of them are precise:
- Manual prompt checks. Ask the same handful of questions in ChatGPT, Gemini and Perplexity on a fixed schedule and log whether your brand appears. Crude, but it is a baseline you can compare against in 30 days.
- Referral traffic from AI platforms in your analytics, filtered by referrer.
- Branded search growth — people looking up a name they met inside an answer.
Two of the eight tasks in any serious GEO plan depend on infrastructure rather than writing: knowing which crawler actually arrived, and knowing how fast you answered it. GPTBot, ClaudeBot and PerplexityBot do not behave like Googlebot. They pull harder, they respect robots.txt differently, and a single slow response during a crawl window can hand the citation to a competitor that answered first.
If your hosting cannot show you per-bot access logs, the content work downstream of it is guesswork. That is not a sales point, it is the measurement layer.
The sequence that actually compounds
If I had to compress all of this into an order of operations for the next 30 days:
| Week | Do this | Done when |
|---|---|---|
| 1 | Rewrite openings of your five most valuable pages as direct answers, 40–60 words | The answer appears before any marketing language |
| 1 | Convert label headings into question-shaped H2s | Every H2 reads as a question or a specific claim |
| 2 | Move comparisons into real HTML tables with header rows and units | The table is readable in source, not an image |
| 2 | Make visible FAQ text match FAQPage JSON-LD exactly | A diff of the two returns nothing |
| 3 | Attach a named author with a role and one verifiable credential | The same author block appears on every page they wrote |
| 3 | Cite one primary source per factual claim | No statistic is unsourced |
| 4 | Run the same prompts across three engines and log the results | You have a dated baseline |
| 4 | Check logs for AI crawler hits and response times | You know whether they reach you, and how fast you replied |
Notice that nothing in weeks one through three requires new content. The fastest GEO gains come from making existing pages extractable, not from publishing more.
Three objections worth answering
"Does this only work for big sites with backlinks?"
No — and small sites often move faster here than in classic SEO. Answer engines weight a clear, self-contained answer heavily. A small site that answers one specific question better than a large competitor can be cited without ranking first anywhere. Backlinks still help discovery; they are not the gate they are in traditional ranking.
"Will GEO replace SEO?"
It builds on top of it. Pages still need to be crawlable and relevant before an engine will consider them at all. The realistic position is both: technical SEO as the floor, GEO tactics as the differentiator.
"Should I block AI crawlers?"
Only if your business model depends on the visit rather than the answer. Blocking GPTBot or ClaudeBot removes you from the pool of citable sources, and with it the referral traffic those citations send. A middle path — allow summarisation, keep proprietary data behind a rule — is usually the better trade.
The one-line version
Before you rewrite another paragraph for the machines, check whether the machines can currently read your page at all. Structure, quotability, schema consistency, measurement. In that order.
Originally published at tanit365.com.
Top comments (0)