GEO, AEO and AI SEO: what’s real and what’s being sold

Generative engine optimisation (GEO), answer engine optimisation (AEO) and AI SEO all describe one job: getting your business named, and your pages cited, when someone asks ChatGPT, Gemini, Perplexity or Google’s AI Mode for a recommendation. Google’s own guide says optimising for its AI features “is optimizing for the search experience, and thus still SEO”.

That matters because a lot of what’s sold under these names in Singapore has no source behind it. The platforms have published what they use. Very little of the sales pitch matches it.

What do GEO, AEO and AI SEO mean?

The term GEO comes from a 2023 paper by researchers at Princeton, IIT Delhi and elsewhere, who tested ways of rewriting content so an AI answer engine would feature it more. Adding quotations and statistics from sources worked best. Keyword stuffing scored below the unmodified page. The headline claim was a visibility boost of up to 40%, measured on a GPT-3.5 test bench, not on live Google or ChatGPT.

AEO is the older marketing term, from the featured-snippet and voice-assistant years. AI SEO is what buyers type into Google. Agencies pick whichever one their market searches for. The work underneath is the same, so when an agency leads with the acronym rather than with what they’ll measure and change, that’s the first thing to notice.

What does Google say it uses?

Google published its guide to generative AI search in May 2026 and updated it in July. Its position:

• AI Overviews and AI Mode are “rooted in” the same ranking and quality systems as ordinary search. There’s no separate layer to optimise for.
• Pages must be crawlable, indexed and eligible to show a snippet. The site also has to be included in generative AI features in Search Console, which is a setting, not a default you can assume.
• “Unique, compelling, and useful” content counts most. Google’s phrase for the opposite is “non-commodity”: if ten other sites say the same thing, there’s nothing to cite you for.
• Structured data helps Google understand a page but isn’t required for AI features.
• The guide lists what it calls myths: no need to chop content into chunks, no need to rewrite it for AI, no need for llms.txt or other AI text files (“Google Search itself doesn’t use them”), and no credit for inauthentic mentions.

Search Console now has a generative AI performance report, so you can see impressions and clicks from AI Overviews and AI Mode separately. Ask for it before you accept anyone’s visibility score.

One figure complicates the “it’s just SEO” line. Ahrefs looked at 4 million URLs cited in AI Overviews in March 2026 and found 38% ranked in Google’s top 10 for the query. In July 2025 the same study found 76%. Ranking still gets you into the pool, but Google is drawing citations from further down the results than it was a year ago. Being first on page one is no longer enough on its own, and being on page three no longer rules you out.

What do ChatGPT and Perplexity cite?

Neither OpenAI nor Perplexity publishes a ranking guide. Both publish crawler documentation, and that’s where the practical answers are.

OpenAI runs separate crawlers. OAI-SearchBot “is used to surface websites in search results in ChatGPT’s search features”. GPTBot collects training data. They’re controlled separately in robots.txt, and OpenAI says plainly that “sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers”. A blanket block on OpenAI, which many security plugins and firewall rules apply by default, takes you out of ChatGPT’s answers while doing nothing to protect your content that a GPTBot block wouldn’t.

Perplexity works the same way. PerplexityBot indexes for search and “is not used to crawl content for AI foundation models”. A second agent, Perplexity-User, fetches a page when a user asks for it, and Perplexity says it “generally ignores robots.txt rules” because a person requested it. Perplexity’s advice for appearing in results is to allow PerplexityBot and its published IP ranges, and to make sure any firewall does too. It names Cloudflare’s Bot Fight Mode as a setting that can challenge its crawler even when you’ve allowed it.

Beyond access, neither company says what makes one page get cited over another. The closest thing to evidence is the Princeton paper: pages that quoted sources and carried specific figures were featured more, and keyword stuffing did worse than leaving the page alone. That was measured on a test bench in 2023, not on today’s ChatGPT, and nobody has published a live equivalent since. Anyone who tells you they know the citation formula for ChatGPT is guessing, including us.

What moved one site’s citations 2.5 times?

A Singapore healthcare group whose WordPress site we maintain went from 40 pages cited by ChatGPT in April 2026 to 103 in October. Google’s AI Overviews went from 46 pages to 118. Perplexity went from 30 to 104. That’s Ahrefs Brand Radar data over six months.

Nothing was sold as GEO. No llms.txt file, no content chunked for AI, no paid mentions, no AI-specific markup. Our work on the site in those six months came out of a security audit, not an SEO plan: a new parent theme went live in late August with the old Bootstrap framework stripped out (about 200KB of unmaintained code gone), PHP moved from 8.1 to 8.4, and security headers tightened. The rest was ordinary support: backups, error fixes, a search bug, an events module. What the site already had was a few hundred service pages that each name a clinic, a procedure and what it involves, on a domain the search crawlers can reach. The citations grew around that.

The platforms were doing the work. Google widened the pool of pages its AI Overviews draw from over the past year (Ahrefs’s 76% to 38% shift is the measure of it), and the ChatGPT and Perplexity numbers on this site moved the same way, though neither company publishes anything that would explain why. A site with specific, crawlable pages gets picked up by that widening without anyone optimising for it. A site with vague pages, or a firewall rule in the way, doesn’t.

What’s sold against what’s published

Sold as GEO or AEO	What the platforms have published
llms.txt file, often as a line item	Google: its search doesn’t use it. OpenAI and Perplexity: no mention in their documentation
Content “chunked” or rewritten for AI	Google lists both as unnecessary
AI-specific schema or markup	Google: structured data not required for AI features
“AI visibility score” from a proprietary dashboard	Search Console reports AI Overviews and AI Mode directly. ChatGPT and Perplexity are tracked by rerunning the same questions, which any tool can do
Guaranteed mentions	Answers change from one run to the next. No platform offers a placement
Paid or arranged mentions on third-party sites	Google names inauthentic mentions as something it discounts
Blocking all AI crawlers “to protect your content”	OpenAI and Perplexity each run a search crawler separate from training. Blocking the search one removes you from answers

The left column is what most Singapore GEO proposals are built from. The right column is why they’re hard to measure afterwards.

How do you check your own visibility in 15 minutes?

Ask the three questions your customers ask. “Best physiotherapist in Tampines”, “who should I use for corporate tax filing in Singapore”, “WordPress agency Singapore recommendations”. Put each one to ChatGPT, Perplexity and Google’s AI Mode. Note who’s named, and which sites the answer links to. Nine answers takes about ten minutes.

Open yoursite.com/robots.txt. Look for OAI-SearchBot, PerplexityBot and Google-Extended. The first two should be allowed. Google-Extended only controls Gemini training and has no effect on Search or AI Overviews, so blocking it is fine. Then read the Sitemap lines. A Singapore restaurant group whose four sites we audited in June had each site’s robots.txt pointing crawlers at the other three brands’ sitemaps, a WordPress multisite misconfiguration, so every crawler visiting the steakhouse was sent off to index the wedding venue. Each sitemap index carried 221 to 436 URLs from the wrong domain. Fixed the same day, confirmed clean a month later.

Then check that your pages are reachable by a link, not only by the sitemap. A metering manufacturer we audited this month had 118 orphan pages, among them news posts still getting 259 visits a month, because its news and insights listings paged by AJAX with no href on the page numbers. A crawler saw nine posts and stopped. The fix was server-rendered pagination links in the theme, done 5 November. Search Console’s Pages report and any crawler (Ahrefs, Screaming Frog) show orphans; the sitemap alone hides them.

In Search Console, open Settings and check that generative AI features are on. Then open Performance and filter by search appearance to see whether AI Overviews and AI Mode are already sending you anything.

If you’re on Cloudflare, open Security, then Settings, and find “AI Crawler and Agent access policies” under Bot traffic. Three settings: Search, Training and Agents. Search must be on “Allow (do not block)”. Block Training if you like; it has no effect on whether you appear in answers. The old single “Block AI Bots” switch is gone, and Cloudflare’s September 2026 notice said zones that had it on were moved to Search allowed and Training blocked, so a blocked Search setting today means somebody chose it. AI Crawl Control, under Security, shows the same thing crawler by crawler. Then check Bot Fight Mode: on the Free plan it’s one switch that can challenge search crawlers with everything else, and on paid plans Super Bot Fight Mode has a “Verified bots” setting that should be set to allow.

The sites cited in those nine answers are your target list. If a directory, review site or publication keeps appearing for your market and you’re not on it, that’s where the outside work goes.

What should you ask an agency?

Which questions will you track, on which platforms, and can I see the baseline before we start? Which of my pages do the answers cite now, and which sites get cited instead of mine? What on my site will you change, and can you show me the line in Google’s or OpenAI’s documentation that says it matters? Who gets the crawler access fixed, you or my developer? What does month one cost and what does cancelling take?

An answer that names a dashboard, an llms.txt file and a content calendar, and can’t point to a published source for any of it, is the sales pitch the table above describes.

We run AI SEO in Singapore as monthly tracking plus the fixes on the site, with a month’s notice to cancel.