GEO in Singapore: how to get cited by AI search when you cannot outrank the incumbents

What GEO actually means

GEO stands for Generative Engine Optimization.

It describes the work of getting your content surfaced and cited inside answers that are generated rather than listed: Google’s AI Overviews, ChatGPT, Perplexity, Claude, and whatever arrives next.

You will also see AEO, Answer Engine Optimization, which is the older label and tends to mean featured snippets and voice results.

AIEO turns up occasionally and has never settled.

The industry has largely converged on GEO, so that is the term we use.

Arguing about the acronym is a waste of a meeting.

The work underneath all three names is the same work.

The important distinction is not the label, it is the mechanism.

Classic SEO competes for a position in a list.

GEO competes to be the source a model reaches for when it writes a paragraph.

Those are different games with different winners, and the second one is far less crowded.

Why you probably cannot outrank the incumbents

We will be blunt, because most agencies will not be.

If you are a newer Singapore business hoping to sit at position one for a broad commercial term like “SEO agency Singapore”, “renovation Singapore” or “corporate lawyer Singapore”, the honest forecast is that you will not get there in the next year or two, and possibly not at all.

The reason is structural, not effort-related.

Those results are held by businesses with five to ten years of accumulated authority and link history, sitting alongside paid directories and government pages.

Search engines lean heavily on age and established trust for exactly these high-stakes commercial queries.

Money does not compress domain age.

Neither does publishing at volume.

Ahrefs put numbers on this by tracking 1.3 million keywords.

Of newly published pages, 1.74% reach Google’s top ten within a year.

Meanwhile 72.9% of the pages already sitting in the top ten are more than three years old, and the average page ranking first is five years old.

In their 2017 run of the same study that last figure was two years.

The door has been closing steadily for a decade.

Read those two numbers together and the position is clear.

You are not competing against a competitor’s content.

You are competing against their calendar.

There is a second squeeze on top of that.

Ahrefs compared 300,000 keywords before and after AI Overviews rolled out in the United States and found the top-ranking page took a 34.5% lower clickthrough rate when an AI Overview sat above it.

That is correlation rather than a controlled experiment, and the study’s own authors say so, but the direction is not seriously disputed.

Position one is worth less than it was, which makes spending years fighting for it a worse trade than it used to be.

The strategic conclusion is not to give up on search.

It is to stop attacking the fortified position and go around it.

Ranking and being cited have come apart

Here is the opening, and it is genuinely new.

Being quoted inside an AI answer and ranking in the top ten of conventional search used to be close to the same thing.

They have come apart, and quickly.

Ahrefs ran 15,000 long-tail queries through ChatGPT, Gemini, Copilot and Perplexity in July 2025 and checked where the cited pages sat in Google.

Only 12% of the cited URLs were in Google’s top ten for that query.

Around 80% did not rank in Google for that query at all.

A separate Ahrefs study looked specifically at Google’s own AI Overviews across 863,000 keyword result pages.

In July 2025, roughly 76% of the pages cited in an AI Overview came from the top ten organic results below it.

By January 2026 that had fallen to 38%.

Google is increasingly answering from pages that are not winning the ranked list at all, because it breaks your question into related sub-questions and cites what it finds across those.

Think about what that means for a business that cannot win the ranked list.

The gatekeeper that kept you out for years, accumulated domain authority, carries less weight in the layer growing fastest.

A model choosing what to quote wants the passage that answers the question cleanly.

It is not checking how long your domain has existed.

Now the caveat, stated plainly so you size this correctly.

Conventional search still sends far more traffic than every AI assistant combined, by a wide margin.

Published multiples for how much better AI-referred traffic converts range from about 1.3 times to figures over twenty times depending on who measured and which industry, which is wide enough that you should distrust any agency quoting you a single confident number for it, including a number that flatters the pitch.

Treat AI citation as an additional channel worth winning early, not a replacement to bet the business on.

How generative engines choose what to quote

Nobody outside these companies has the ranking logic, and anyone who tells you they do is selling something.

What we can describe is the observable behaviour, which is consistent enough across engines to plan around.

A generative engine is assembling an answer, then attaching sources to the claims in it.

So it favours content that is easy to lift a defensible sentence from.

In practice that means:

• A direct answer, early. If the answer to the question is in paragraph one, it can be quoted. If it arrives in paragraph nine after a preamble about how the digital landscape is changing, it will not be.
• Structure that matches the question. Headings phrased the way people actually ask, followed immediately by the answer to that heading.
• Specifics over adjectives. A number, a date, a named source, a real constraint. “Fast turnaround” is unquotable. “Most audits take us nine to twelve working days” is quotable.
• Clear attribution. An identifiable organisation or author behind the claim, consistent across the site and matching what other sites say about you.
• Information that exists nowhere else. This is the strongest lever by a distance, and it gets its own section below.

Notice that none of this is a trick.

It is the same content that a human reader would call useful.

That is deliberate on the engines’ part, and it is why GEO work does not decay the way tactical SEO does.

The work, in order

If we were handed a site tomorrow and asked to make it citable, this is the sequence, and the order matters more than the list.

1. Fix the foundations first

If pages are slow, blocked from crawling, duplicated across near-identical location pages, or thin, stop here and fix that.

Most of that hurts both ways, because a search engine and a generative one both need to fetch, parse and trust the page, and there is no GEO tactic that routes around a broken foundation.

Most of it is also the SEO work you needed anyway.

Two exceptions, and they are the reason we stopped saying the two jobs are simply the same job.

JavaScript is the one that catches people out.

Google renders it.

Google’s own documentation says “Google Search runs JavaScript with an evergreen version of Chromium” and that “a headless Chromium renders the page and executes the JavaScript”.

The AI crawlers appear not to.

We say appear deliberately: not one of OpenAI, Anthropic or Perplexity states this either way on its own crawler documentation, and we read all three this month.

The best evidence that exists is a study Vercel published with MERJ in December 2024, which observed the crawlers requesting JavaScript files and never running them, at 11.50% of ChatGPT’s requests and 23.84% of Claude’s.

That is one study, roughly twenty months old by the time you are reading this, and it states no sample size.

Best available answer, not a settled fact.

We would rather tell you the evidence is thin than let you assume it is not.

If it holds, the consequence is specific and uncomfortable.

A site that assembles its content in the browser can pass a conventional SEO audit, rank in Google, and still hand a generative engine an empty page.

Server-rendering or pre-rendering to fix that is real extra work, and no ordinary SEO checklist will raise it, because Google was never the one struggling to read you.

“Blocked from crawling” is not one switch either.

OpenAI alone documents four tokens doing different jobs: GPTBot for model training, OAI-SearchBot to “surface websites in search results in ChatGPT’s search features”, ChatGPT-User for fetches a person triggers, and OAI-AdsBot for ad safety checks.

A business that blocked GPTBot to stay out of model training, which was a reasonable call and a common one, has not blocked the bot that decides whether ChatGPT can cite it.

Different lines in robots.txt.

Check which one you actually blocked before concluding the engines are ignoring you.

Google’s version of this is the mirror image.

Blocking Google-Extended achieves nothing here. Google states it “does not impact a site’s inclusion in Google Search nor is it used as a ranking signal in Google Search”. It has no crawler of its own. The token is “used in a control capacity” over Gemini training.

AI Overviews are served from the ordinary index, so the directive that governs them is the Googlebot one, which is the same directive governing whether you appear in Search at all.

2. Put the answer at the top of the page

Take your twenty most valuable buyer questions and make sure each has a page, or a clearly-headed section, that answers it near the top.

Then elaborate underneath for the human who wants depth.

We should be straight about the standing of that advice, because it is the most confidently repeated instruction in this field and the evidence for it is thinner than the confidence.

No AI provider documents how it splits a page up for retrieval, so nobody outside those companies knows what “near the top” buys you.

Google, meanwhile, says the opposite of what the advice implies: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.”

Its passage ranking was built specifically to surface answers that are not near the top.

We still do it, and we still recommend it, for reasons that survive without a study: it follows from how a model assembles a quotation, it costs nothing, and it plainly serves a human in a hurry.

That is a working heuristic, not a measured lever.

Anyone quoting you a specific word count for it is quoting a convention.

3. Publish something only you have

This is the part most businesses skip, and it is the part that compounds.

A model cannot summarise a figure it can only get from you without pointing at you.

Neither can a journalist or a competitor’s blog post.

Original data does not require a research department.

It requires counting something you already touch.

A property firm has transaction records.

A clinic knows its real waiting times.

A logistics company knows what a route actually costs this quarter.

We built a Singapore pool guide on official ActiveSG data: twenty-eight pages covering all twenty-five complexes in ActiveSG’s public directory, plus three that have closed or have not opened yet, each carrying its current status.

It earns search traffic with no advertising spend behind it.

That is the shape of the play: take a subject where the public information is scattered, assemble the complete version, and keep the status current.

A directory that still lists a pool which has shut is worth less than one that says so, which is the whole of step 6 below in a single example.

Note what that costs.

The complete version is the easy half.

Staying correct is the half that keeps paying, and it is the half most published “ultimate guides” quietly stop doing in month four.

4. Get named on pages you do not own

Generative engines lean on third-party corroboration.

If the only place claiming you are a specialist in something is your own homepage, that claim is weak.

If it also appears in an industry publication, a conference speaker page, a genuine directory listing and a forum thread where somebody recommended you unprompted, it holds.

In practice this means the unglamorous work: complete and consistent business listings, real reviews, contributing writing to publications your buyers read, showing up where your market actually discusses this.

One caution, because this is exactly where well-meant outreach becomes a policy problem.

Google’s spam policies name as link spam “advertorials or native advertising where payment is received for articles that include links that pass ranking credit, or links with optimized anchor text in articles, guest posts, or press releases distributed on other sites”.

The fix is not to avoid the work, it is to label it.

Google says plainly that such links are fine “as long as they are qualified with a rel=”nofollow” or rel=”sponsored” attribute value”.

An unpaid editorial mention needs nothing.

Anything you paid for, or got a free product or service for, needs the attribute.

The same policy lists “low-quality directory or bookmark site links” separately, which is why the word genuine is doing real work above.

On consistency we will say only what is documented.

Keeping the same business name, address format and phone number everywhere is long-standing practice, it is what we do, and it plainly helps any system work out that a pile of listings describe one business.

But Google’s own local ranking guidance does not mention it.

That page names three factors and only three: relevance, “how well a Business Profile matches what someone is searching for”; distance, “how far each business is from the customer who’s searching”; and prominence, “how well-known a business is”.

Treat consistency as sensible hygiene rather than a documented ranking lever, and be wary of anyone selling it as the latter.

5. Ship the machine-readable layer

Structured data is how you tell a parser what a page is rather than making it guess.

Organization and LocalBusiness markup for the business itself, Article for guides, Breadcrumb for hierarchy.

Each of those has its own page in Google’s structured data reference and can be checked in Google’s Rich Results Test.

It is cheap, and in our experience most competitors have gaps in it.

Worth knowing, because it is widely recommended and we have recommended it ourselves: Service markup is not one of these.

It is real schema.org vocabulary and it describes what you sell perfectly well, but Google does not list it as a supported type and has never published a reference page for it.

The address where that page would sit returns a 404.

Publishing it costs nothing and harms nothing, so keep it if you have it.

Just do not let anyone bill it as something Google acts on, and note that the Rich Results Test cannot confirm it either, because there is no Google feature for it to confirm against.

Do not bother with FAQPage.

Google has retired that rich result and taken the documentation down with it: the old page now redirects to a changelog entry saying the feature is no longer shown in search results.

Keep the question block on the page if it answers something real, because a good answer written plainly is still the thing a model quotes.

Just do not expect the markup to earn you anything.

Now llms.txt, a file that is supposed to describe your site to AI crawlers, and which a lot of agencies are currently selling.

Google’s own documentation for site owners is unusually direct about it: “Google Search itself doesn’t use them”, and creating one “will neither harm nor help your site’s visibility or rankings in Google Search, as Google Search ignores them”.

No other major AI provider has publicly confirmed that it reads publishers’ llms.txt when deciding what to cite either.

We do not publish an llms.txt file on this site.

We would not put it on an invoice.

If an agency is presenting llms.txt as a significant part of what you are paying for, that is a useful signal about the rest of the proposal.

6. Keep it current, visibly

Conventional search favours content that looks maintained, and generative engines appear to behave the same way, though Google is the only one that documents anything here.

Dates that are real, figures that match the current year, links that still resolve.

A guide last touched in 2024 signals abandonment.

Reviewing your top twenty pages quarterly beats publishing forty new ones a year that nobody revisits.

The load-bearing word in that paragraph is real.

Google’s helpful-content guidance lists, among the questions it tells you to assess your own site against, “Are you changing the date of pages to make them seem fresh when the content has not substantially changed?”

So a quarterly review has to produce an actual change to the page, not a new date on an unchanged one.

If nothing needed changing, the honest move is to leave the date where it is.

This is the failure mode of the advice, not an argument against it: the review is the point, the date is only its receipt.

The Singapore layer

Three things are specific to competing here.

The local results are the faster lane.

For anything with local intent, and in Singapore that is most service businesses, the block of map results is winnable in a fraction of the time a head term takes.

Worth being precise about what actually moves it, because this corner of the industry is thick with confident advice.

Google states that “local results are mainly based on relevance, distance, and popularity”, and that prominence is “also based on info like how many websites link to your business and how many reviews you have.

More reviews and positive ratings can help your business’s local ranking”.

So: a complete and accurate profile, links, review count and rating.

What that page does not mention, anywhere, is review velocity, citation consistency or publishing location-specific content, all three of which get sold as local ranking factors.

They may still be worth doing.

They are not documented.

If you need leads this year rather than in 2029, this is where to spend first.

Understand the listicle market before you trust it.

Some of the “best agency in Singapore” style lists that rank on page one are published by companies that appear on those lists themselves, in first place.

We will not tell you it is most of them, because we have not counted and nobody else has published a count either.

We will tell you that we ran the check while writing this and the first example took under a minute to find: a “10 best SEO agencies in Singapore” list for this year, sitting on an agency’s own domain, with that same agency at number one and no disclosure on the page.

Some publishers do disclose the arrangement and state a methodology, which is fair enough.

Before you take any such ranking seriously, or pay to be added to one, check whose domain the list sits on and whether the top entry shares an address, phone number or owner with the publisher.

That check takes two minutes and it changes how the whole page reads.

We would rather teach you to run it than hand you our own list to trust.

Language is an unclaimed gap, but not the one it is usually sold as.

Very few competitors here run a genuinely bilingual site rather than a machine-translated afterthought, so the gap itself is real.

The population data is more interesting than the pitch that normally accompanies it.

The Department of Statistics reported in its General Household Survey 2025 release, dated 30 June 2026, that English is “the most frequently spoken language in Singapore homes, used for nearly six in ten residents (58.1%) aged 5 years and over, up from 48.3% in 2020”.

Read only that and a second language looks like a shrinking bet.

The next sentence complicates it: of those English-first residents, “the majority (80.5%) also spoke a vernacular language as a second language”, and about nine in ten literate residents under 45 can read two or more languages.

So the opportunity is not a monolingual Chinese or Malay audience sitting unserved.

It is a bilingual one that can read either version and will stay with whichever actually answers the question.

That is a weaker argument for translating everything and a stronger one for writing your best pages properly in a second language.

One honest limit on all of the above: none of it measures the language people search in.

We could not find a published source that does, from the Department of Statistics or anyone else.

So if an agency quotes you a percentage of Singapore searches conducted in Chinese, ask where the number came from, because we could not find one to quote you.

What we would not bother with

Equally useful, and rarer to find written down.

• Chasing head terms with content volume. Publishing two hundred thin posts to catch a term held by a ten-year-old competitor burns budget, and the harm is documented rather than folklore. Google’s spam policies name scaled content abuse, “when many pages are generated for the primary purpose of manipulating search rankings and not helping users”, adding that this applies “no matter how it’s created”, so writing them by hand is no defence. The stated consequence: “Sites that violate our policies may rank lower in results or not appear in results at all”.
• Paying for listicle placement. You are renting a position on a page whose owner ranks themselves first. The moment you stop paying, it is gone, and it builds nothing you keep. There is a second cost nobody selling you the slot will raise: if the paid placement carries an ordinary followed link back to you, that is Google’s own worked example of link spam, “exchanging money for links, or posts that contain links”, and the risk lands on your site, not the publisher’s. If you buy one anyway, insist the link is marked sponsored. That keeps it inside the rules and, tellingly, removes most of what you were sold.
• Keyword density and similar folklore. This has not been how ranking works for a long time.
• Any promise of guaranteed AI citation. There is no submission form, and no way to buy the recommendation itself. Be precise about this one, because it moved in 2026. Google’s own advertising documentation now states that text and shopping ads are “eligible to show within the AI Overviews”, and the other major assistants have been introducing labelled ad slots of their own. What none of them sells is a place in the sentence the model writes, and Google’s spam policy now names “attempting to manipulate generative AI responses in Google Search” as spam outright. Buying an ad slot next to an answer is a normal media buy. An agency guaranteeing you a citation is either misinformed or hoping you are.
• Rebranding your existing SEO retainer as GEO at a higher price. Ask precisely what changes in the work. If the answer is vague, the answer is nothing.

How to measure it

GEO has a real measurement problem, though it is smaller than it was a year ago.

Google added a generative AI performance report to Search Console in June 2026, covering AI Overviews and AI Mode.

Turn it on if you have it, and read it knowing two limits: it reports impressions only, so no clicks and no queries, and Google is still releasing it to a subset of site owners, so you may not have it yet.

In the ordinary Performance report those same AI features are counted inside the standard Web search type, which is Google’s own documented behaviour, so do not expect the headline numbers there to tell you anything about AI separately.

For ChatGPT, Perplexity and the rest, nothing of the kind exists for site owners at all.

So these three imperfect measures still carry most of the weight.

Referral traffic from assistant domains.

Google Analytics now has a built-in AI Assistants channel, so this needs no setup: arrivals from ChatGPT, Gemini, Copilot and similar land in it automatically.

Read the definition before you trust the number, because there is a trap in it.

Google’s own AI Overviews and AI Mode are deliberately excluded from that channel and counted under Organic Search instead.

So the AI Assistants figure is every assistant except Google’s, and anyone reading it as their total AI traffic is understating it by the largest source.

Watch the trend and the conversion rate, not the raw count, which will look small for a long time.

A fixed prompt set, run monthly.

Write down the twenty questions a buyer would genuinely ask before hiring you.

Run them through the major assistants on the same day each month and record whether you are named, how you are described, and who is named instead.

It is manual and it is the most honest signal available.

Keep the wording identical between runs or the comparison is meaningless.

Branded search volume.

Being quoted often does not produce an immediate click.

It produces somebody searching your company name a week later.

A rising branded search trend alongside flat conventional rankings is frequently the AI layer working.

Set the expectation honestly with whoever signs off the budget: this moves in months, and the early numbers will be small.

Questions we get asked

Sources and method

This guide was written from the published research on AI citation behaviour, Google’s own documentation for site owners, and a scan of how the Singapore agency market presents itself.

Sourcing runs two ways here.

The statistics in the box above name the organisation that published them, link to the original, and state the sample and the date.

Quotations from documentation, which is now most of the operational advice, name the publisher in the sentence and are reproduced word for word rather than paraphrased, so you can search the quoted phrase and land on the source yourself.

Where the only evidence available is a single study, or a single example we found ourselves, the sentence says so rather than rounding it up into a consensus.

What changed on 1 August 2026.

Audited the operational advice against current documentation.

Added the JavaScript rendering gap between Google and the AI crawlers, the fact that blocking GPTBot does not block the bot that cites you, the link-marking rule for anything paid, and Singapore language data from the June 2026 General Household Survey release.

Downgraded the answer-in-the-first-hundred-words rule to the heuristic it actually is, and corrected three local-ranking factors we had repeated without checking.

We started with a longer list of statistics and checked each one against its original source before writing.

Several did not survive, and it is worth naming them, because you will see them quoted elsewhere.

A widely-shared claim that Google sends 345 times more traffic than the AI assistants combined comes from a March 2025 snapshot that the same publisher had already revised down to 210 times by that June, so we left the multiple out entirely.

A figure of 46.5% for AI citations ranking outside Google’s top 50 could not be traced to any primary source at all.

Neither could a tidy claim that AI traffic converts 2.4 times better, where the real published range runs from roughly 1.3 times to over twenty.

One commonly-cited range for pages reaching the top ten turned out to splice together two different years of the same study.

Four numbers that would have made this guide look more authoritative, removed for the reason that we could not stand behind them.

Three limits on what remains.

Search results vary by location, and rank order inside google.com.sg specifically cannot be verified from published research alone, so nothing here claims a Singapore-specific ranking.

The AI Overview clickthrough figure is correlational, not a controlled experiment.

And AI citation behaviour is moving fast enough that the top-ten overlap number changed by half in six months, so treat every figure here as dated rather than permanent.

We date our guides for that reason and update them in place with a note on what changed.

If you spot something here that is wrong or out of date, tell us and we will correct it with a note saying what changed.