Singapore

How to optimise for AI search: a measurable operating method for marketing teams

As of November 2026, the practical answer for a Singapore marketing team is four steps: fix a question set, measure how often your brand is mentioned and how often your own site appears as a source, refresh the pages engines already cite instead of writing new ones, and put the result on one page you can defend in a budget meeting. Anyone shopping for generative engine optimization singapore work needs that page before the first invoice. Hootsuite published its own practical guide to the topic on 12 November 2026, and this article takes the same question and answers it at the level where the money is decided.

Key takeaways

Ahrefs measured the presence of an AI Overview as correlating with a 58% lower average clickthrough rate for the top-ranking page, up from the 34.5% fall the same team recorded eight months earlier.
Recency in AI citations is manufactured by updates, not by new publishing: 72% of cited pages look fresh by last update, but the share first published in the last year falls well below that.
Third-party pages carry almost all the weight: about 98% of what gets cited is not yours.
Twenty to 30 real buyer questions, repeated every month, is enough to show a trend without a research budget, and that is the level a generative engine optimization singapore brief should be held to.

Why organic clicks fall before rankings do

The click loss is measured at the top of the page, which is where most brands never look. Ahrefs built its latest comparison on a keyword sample split evenly between informational queries with an AI Overview and informational queries without one, 300,000 keywords in all, half on each side. Some AI Overviews sit outside position one, so a page can lose clicks without any visible change to its own snippet.

Two readings of the same phenomenon, eight months apart, matter to a planner. The earlier study measured AI Overviews reducing clicks to top-ranking content by 34.5%. The later one widened the correlation to 58%. The gap between the two numbers is the point: the effect grew, and a forecast built on the earlier figure understates what a content budget is now buying. Ahrefs is careful to call this a correlation, and that framing is correct. A page with an AI Overview above it differs from one without it in more ways than the box itself.

A marketing manager leans towards a laptop where a summary panel sits above the first ordinary search result.

How AI engines decide what to cite

The useful part of the research is not that engines like fresh content; everyone says that. It is which kind of fresh.

Seer Interactive tracked four brands across three engines over a single four-month window, then dated the pages they were cited on. The dated set holds 7,683 pages. Its July reading: 75% of cited pages had been updated in the last year and the great majority within two. The pages cited in all four months, which the study calls the always-on set, are 45% of the dated set, with most of them updated in the last year and a median update age around six months. That median is the closest thing to a cadence rule the data offers.

The refresh finding, in one pair of numbers

Take the cited pages where both dates are readable. By last update, 72% look fresh. Measured by publish date within the last year, the share drops. Among the highest-volume cited pages the pattern is sharper: 75% were updated in the last year, but only a minority were first published that recently.

The implication for a content plan is blunt. A page you refreshed two years after writing beats a new page you published last month, and the budget line should sit on maintenance rather than on volume. Half of cited pages are blogs and guides, and most of those were updated in the last year, which is where the refresh work earns most.

Engines are not interchangeable

ChatGPT leans on editorial: blogs and guides take 46% of its citations, and most of the content it cites was updated in the last year. Gemini is the freshest of the three, with almost all citations within two years. Perplexity is the forgiving one. A team that reports one blended freshness figure across three engines is describing an average of three different rules.

Where third-party pages carry the answer

Searchable's September reading of cited social platforms covers a month of answers across six engines. YouTube appeared in 31.2% of AI Overviews answers and Reddit in 17.7%. Seer's finding sits underneath that: you may own a tiny share of what gets cited, and the other 98% is third-party.

The Reddit numbers are worth a second look because they kill an easy assumption. Of a random sample of cited threads, two in three had fewer than 20 upvotes and the typical one had 9; the median thread was 635 days old and the median text post short. When a cited page was the brand's own site, the engine named the brand most of the time. Searchable found that when ChatGPT named one of the brands it tracks, it also cited that brand's own website 71% of the time; on Gemini that fell sharply below that, so more often than not the description came from someone else's page.

How to optimize for AI search with measurable definitions and benchmark numbers

This is the gap the search results leave open: every agency page explains what GEO is, none hands over a method with a denominator. Here is one you can run in a quarter, written so that a finance team can audit it.

Step 1: fix the question set and its version

Write 20 to 30 real buyer questions, the ones your sales team hears, not the ones your brand wants, and lock the wording. Twenty to 30 is enough for a monthly trend and small enough to run by hand; that minimum is the one Searchable recommends. Version the list. A change to wording is a new version, and a comparison across versions is not a comparison.

Step 2: compute two rates, never one score

Mention rate is the number of valid answers containing your brand divided by the number of valid answers under identical conditions. Source-appearance rate is the number of valid answers in which your own site appears, divided by that same valid answer group. Publish numerator, denominator and sample size together, per engine. A single "visibility score" hides which of the two moved, and they move for different reasons. Not being mentioned is not the same as being described negatively; keep those two counts apart in the report.

Step 3: use retrieval data to explain the gap

When a question returns no mention of your brand, there are three separate cases, and they call for different work. Your site is not among the retrieval sources, which points at crawlability and page structure. Your site is retrieved but not visibly cited, which points at content that lacks a direct answer or verifiable facts. Your site is cited but the description is wrong, which points at outdated brand facts. Retrieval counted here is not the same as visits; referral traffic and conversions need your analytics, and nobody should read them off a citation table.

Step 4: refresh on the six-month rule and re-test on the same version

Take the median update age of around six months from the always-on set as a starting threshold, not a law. Flag owned pages cited in the last year and older than that median; refresh the ones that carry buyer questions. Re-run the same question version and report the two rates side by side. If the question set changed, start a new baseline.

What this method will not tell you

It will not tell you what any individual user sees, because sampling runs through an engine's API. It will not tell you that a rise caused a sale; a mention rate moving in the same quarter as a lead count is a correlation until you can separate the two. It also will not survive a provider that reports a change without the question version and the sample size attached.

An analyst points at two tallied columns on a whiteboard while a colleague holding a printed sheet asks a question.

When a GEO agency in Singapore earns its fee

The decision most teams actually face is not "should we do AI search" but "do we hire for it". Use the method above to decide, and use three tests on the agency. Teams that buy generative engine optimization singapore work without those tests usually end up paying for activity reports instead of answers.

The three tests

Can it state its mention-rate and source-appearance-rate definitions, with numerator and denominator, in writing before the engagement starts?
Does it separate retrieval from citation, and say so when it reports on either?
Does its content plan spend most of its effort on pages that already get cited, rather than on new volume?

A geo agency in singapore engagement should also state what it does not promise. No provider controls what an engine recommends, and a proposal that implies otherwise has told you something about how it will report.

What an engagement should contain

The scope should list the engines, the question count, the competitor set, the content volume and who publishes. Content delivery, client approval, publishing, indexing and citation are five statuses, and a report that collapses them into one is not auditable. Where the engagement covers content, each piece should answer one buyer question with a direct answer, the necessary background and verifiable facts, because that is the shape a citation can be built from.

How Topkee approaches this

Topkee runs a Generative Engine Optimization service built around the same three moves a reader can check: ASK, UNDERSTAND, IMPROVE. ASK starts from real buyer questions rather than brand-name tracking; UNDERSTAND covers prompt tracking, query fan-out, sentiment and description, competitor comparison and AI browsing analysis; IMPROVE closes the gaps with knowledge content writing. AI visibility measurement in this design is defined, not asserted: mention rate is valid answers containing the brand divided by valid answers under identical conditions, and source-appearance rate is valid answers containing the brand's own site divided by the same group, with question version, engine, language, region and sampling cadence fixed before any comparison. Query fan-out reads the sub-queries an engine actually returns, so it does not claim a complete retrieval path when data is missing. AI browsing analysis separates a page being retrieved from being visibly cited, and it is not a visit count. GEO Insight shows brand ranking, visibility, competitors and citation sources across a fixed number of active tracked questions that depends on the plan; the monitoring samples come from an API environment and are not what each consumer sees. Topkee also applies E-E-A-T and SEO copywriting discipline to the refresh work, and publishes no guarantee of AI recommendation, fixed visibility or query volume. The service is delivered from Topkee Pte. Ltd., the Singapore operating company established in the final quarter of 2024.

A monthly scorecard a finance team can read

One page, a handful of fields, no dashboard required.

Field	What to write	Why
Question version	v1, dated, with the count	A change here resets every comparison
Engines	Name each engine, separately	Gemini and Perplexity follow different rules
Valid answers	Per engine, with failures counted	The denominator nobody checks
Mention rate	Numerator / denominator	The outcome the brand cares about first
Source-appearance rate	Numerator / denominator	Separates "we are known" from "we are the source"
Actions open	Recommended, approved, implemented	Keeps a promise out of a result column

The disclosure fields are the whole product, and they are what separates a geo agency in singapore report from a slide deck. A firm quoting a brand mention rate without the question version, the engine list and the valid-answer count is quoting a number you cannot re-run.

A team leader holds up one printed page of figures in a budget meeting while a finance colleague takes notes.

Where to start next quarter

Build the question list in the next two weeks; you need 20 to 30 questions and a spreadsheet. Then run one baseline round on two engines and publish the two rates internally with their definitions. [AEO vs. Search Rankings: AI Visibility Basics](https://www.topkee.com.sg/dynamic/ai-visibility-vs-search-rankings-is-your-agency-measuring) sets out the difference between a ranking and an AI mention if that distinction is still being argued inside your own team. If the baseline shows your site is retrieved but never cited, fix the top three pages before you hire anyone, and only then look at a geo agency in singapore shortlist. If it shows your site is not retrieved at all, start with crawlability, and read how closed-loop measurement connects AI answers to pipeline before you build a reporting layer you will have to rebuild.

FAQ

How long before a refresh shows up in AI answers?

The Seer data gives a median update age of around six months among always-on cited pages, which says what cited pages look like, not how fast a change appears. Measure monthly on a fixed question version and judge from your own trend.

Is AI search optimisation different from SEO?

Partly. The content is the same asset, but the outcome is a mention and a citation rather than a rank, and the denominator is a fixed question set. SEO copywriting and E-E-A-T work still apply; the reporting does not transfer.

Should we hire a GEO agency or do it in-house?

If you can run 20 to 30 questions monthly and own a content editor, in-house is feasible at the baseline stage. Hire when you need third-party placements, because about 98% of citations come from pages you do not own.

Do we need to beat our competitors in every engine?

No. Engines differ: Gemini leans hardest on recently updated content, Perplexity least. Pick the engines your buyers use and compare within each, on the same question version.

What is a realistic first-quarter target?

A defensible target is a measured baseline plus a documented refresh cycle, not a visibility percentage. Any provider offering a guaranteed mention rate is promising something no one controls.