Generative engine optimisation (GEO), what actually works in 2026

GEO. AEO. AI SEO. LLM SEO. Answer engine optimisation. Generative search optimisation. The acronyms keep piling up and the work underneath them barely changes. This post strips out the jargon and gives you five moves that actually shift AI citations for a small business. Nothing theoretical.

What GEO really means

Generative engine optimisation is making it easier for AI models to quote your site. That is it. The "engines" in the phrase are ChatGPT, Perplexity, Claude, Gemini, and whatever launches next month. They all need the same things to cite you cleanly.

AEO (answer engine optimisation) is the older cousin. Same work, coined when the answer engines were Google featured snippets and Alexa. The techniques port across almost without change.

Ignore the label debate. Do the work.

Move one: rewrite your homepage first 40 words

Open your homepage. Read the first 40 words aloud. Do they answer "who is this, what do they do, where, and why should I trust them"? If they start with "Welcome to..." or "We are a family-run..." or a brand story, the model has nothing to quote.

A working first paragraph looks like this:

"Hove Dental Care is a private and NHS dental practice in Hove, East Sussex. We offer emergency appointments within 24 hours, general dentistry, and Invisalign. Rated 4.9 from 312 Google reviews, open Monday to Saturday."

Who, what, where, proof. Forty words. That paragraph is what a model will lift verbatim when someone asks "emergency dentist Hove".

Move two: add FAQPage schema to your top three pages

FAQPage schema is a JSON-LD block that tells search engines and AI models, "these are the questions this page answers". When a user asks one of those questions to ChatGPT, the model can match directly.

Pick the top three pages by traffic or commercial intent. For each one, write five to eight real customer questions and plain-English answers. Paste the JSON-LD into the head of the page. Validate with Google's Rich Results Test.

Questions to avoid: "Why choose us?", "What makes us different?". These are brand questions, not customer questions. A real customer question is "how much does a crown cost?" or "do you take same-day emergency bookings?".

If writing schema by hand feels fiddly, the free schema pack has it pre-filled for five trades.

Move three: LocalBusiness schema with full NAP

If you serve a geography, LocalBusiness schema is non-negotiable. Name, address, phone (the full NAP), opening hours, geo coordinates, areas served, and the sameAs array linking to your Google Business Profile, Facebook, and LinkedIn.

This is the block that lets a model confidently say "yes, this business is in Leeds, yes it's a real one, here is the phone number". Without it, the model hedges or picks a competitor with cleaner data.

Common failures I see on client sites:

• Address missing the postcode.
• Phone number formatted differently in the schema than on the page.
• No sameAs links, so the model cannot cross-check identity.
• Opening hours listed as text but not in the schema block.

Fix those and you are ahead of most local competitors.

Move four: AggregateRating schema with a real review count

AI models weight businesses with visible social proof. AggregateRating schema publishes your review count and average rating in a way the model can parse. It pulls from your Google Business Profile or a connected review platform such as Trustpilot or Reviews.io.

One caveat, and this is where the FTC rule bites. Do not make up reviews. Do not generate them with AI. The Consumer Reviews and Testimonials Rule in effect since October 2024 penalises fake reviews at $51,744 per violation. If you have 14 real reviews, publish 14. Work on getting to 50 over the next 90 days by asking customers after delivery.

Real reviews, honestly displayed, beat inflated numbers every time. Models cross-check against Google and catch discrepancies.

Move five: monthly citation monitoring

You cannot improve what you do not measure. Once a month, run ten customer prompts through ChatGPT, Perplexity, Claude, and Gemini. Write down which domains each one cites. Keep a simple spreadsheet.

Pick prompts that match real buying intent:

• "best [your service] in [your town]"
• "[your service] near me"
• "who is the most reviewed [your trade] in [your city]"
• "cheapest [your product] UK"
• "[your specific niche term] recommendations"

If you are not in the top three cited domains on a prompt after 90 days of work, the issue is usually one of three things, cached model, missing schema, or thin off-site signals. I wrote a separate post on diagnosing stuck citations.

What to ignore

A few things being sold as GEO that are not worth your time right now:

• Llms.txt. Nice idea, no major model uses it as a ranking signal in 2026. Add it if you like, expect nothing.
• AI content at scale. Writing 400 thin pages with an AI tool hurts more than it helps. Models can now detect generic AI prose and discount it.
• Paid "AI visibility boost" services that promise guaranteed ChatGPT placement. No one can guarantee it. Walk away.

A realistic timeline

For a single-location small business doing the five moves above:

• Week 1: homepage rewrite, LocalBusiness schema live.
• Week 2: FAQPage schema on top three pages.
• Week 3: review system in place, AggregateRating wired up.
• Week 4: first citation monitor baseline.
• Weeks 5 to 12: consistent publishing, review flywheel, monthly monitor.

Most clients see the first new ChatGPT citation somewhere between week 3 and week 8. The models refresh their indices on their own schedules, so results are lumpy, not linear.

Where to go next

If you want the exact schema blocks, the free snippet pack has them filled in for five trades. If you want a priority list for your own site, the audit gives you one with the copy and schema written.

Happy to answer anything, just reply.

Bob