Singapore

AI Knows Your Brand But Rarely Shortlists You

Ask an assistant about your own company by name and it will almost certainly describe you correctly. Ask it the question a buyer actually types, and you may not appear at all. That distance between being known and being shortlisted is the part of aeo, answer engine optimisation, most teams measure last, and it is the part that decides who reaches a request for proposal.

Before the 2027 planning cycle starts, a marketing lead at a Singapore professional-services or B2B supplier has one workable decision to make: keep reporting rankings and branded-query answers, or add a recommendation rate measured from real buyer prompts, repair the sources behind a stale description, and tie the result to pipeline. This analysis argues for the second and gives the steps to run it.

As of November 2026, the evidence points the same way from three directions: buyer behaviour has moved, the brand descriptions engines produce lag behind reality, and the cost of that lag now shows up outside the marketing report.

Key takeaways

Being known and being recommended are two different measurements, and a branded query tests only the first.
A small, fixed prompt set run on a schedule tells you what a single test never can: whether the list you are missing from is moving.
Answer engines describe the company you used to be, because they draw on whatever the web said most often.
Repairing the cited record is cheaper than earning a place on the roundups engines cite, and it should come first.
The value of a shortlist shows up in revenue per employee and in what an acquirer will pay, not in session counts.

Why being known is not being recommended
Building a prompt set that tests recommendation
Repairing the sources behind a stale description
Shortlists are built from other people's lists
Connecting a shortlist to signups and pipeline
What Topkee measures, and what it will not promise
FAQ

Why being known is not being recommended

Tony Uphoff, who now writes the Uphoff on Media newsletter, ran a test worth copying. He put the questions his prospective clients ask to ChatGPT, Gemini, Perplexity and Google AI Overviews, then counted how often each engine named the two businesses he started. Across 59 answers, his newsletter appeared in none of the category answers. His advisory firm appeared in one of 43. The same engines, asked who he is, described his career without hesitating.

He summarises the result plainly: the models know the person and almost never recommend what he has built. For a marketing lead, that is the false positive to look for in your own reporting. A branded query returns a confident, accurate paragraph, the dashboard shows green, and nobody notices that the brand is absent from the comparison question that matters.

The cause is structural rather than technical. A named query asks an engine to retrieve what it already holds about an entity. A category query asks it to compose a shortlist from everything it knows about a market, which means the brand now competes on evidence it may never have published: comparison pages, third-party roundups, review profiles and clear statements of who the service is for.

The buyer side has moved in the same direction. G2's survey of B2B software buyers, run in March this year, found 51% now start research with an AI chatbot more often than with Google, up from 29% in its 2025 report, and 71% rely on AI chatbots at some point in the process. The starting line has moved, so a scorecard built around search positions is measuring a smaller share of the decision than it did two years ago.

A buyer types a question into a chat window on a laptop, a handwritten shortlist open beside the keyboard.

Measuring your AI recommendation rate with a fixed prompt set

The measurement is not hard; it is just unfamiliar. What the industry calls aeo work only becomes measurable once the questions are fixed, and the news report's advice is to write 10 to 15 buyer prompts in the buyer's own words, built around the category, the problem, and "who should I hire for", then add two or three branded prompts as a control. Run the set weekly rather than once, because an answer engine's output varies between runs and a single week cannot tell a trend from noise.

A prompt set for a Singapore firm can be assembled in an afternoon. It needs four groups, and the groups should not be mixed when you report them.

Prompt group	Example wording	What it tests
Category	"best corporate training providers in Singapore for mid-sized firms"	Whether the brand appears in a generic recommendation
Problem	"our finance team spends three days on month-end close, what should we change"	Whether the brand is named against a specific pain
Selection	"who should I hire to redesign a warehouse layout in Jurong"	Whether the brand appears in a hiring question
Branded (control)	"what does <brand> do"	Whether the description is accurate, not whether the brand is recommended

Most aeo reporting fails on comparability rather than on data volume, so four habits matter. Fix the engine, the model mode, the language and the region before you start, and keep them fixed. Keep branded and category samples in separate reports, since one measures description quality and the other measures presence. Score description accuracy separately from mention, because being named with an outdated role or a wrong service scope is a different repair from not being named at all. Record absence as absence; a brand missing from a comparison answer has a data gap to investigate, not a reputation problem to defend.

The same reporting logic applies whether you run the prompts by hand or through a service. Mention rate is calculated as valid answers containing the brand divided by all valid answers. It is worth naming the difference out loud, because it is where most reporting breaks: a mention rate is not a shortlist rate. Being mentioned, being described accurately, being cited as a source, and being listed as a recommended option sit at four different levels, and only the first three can be measured from answer text alone.

An analyst compares four stacked columns on a wall chart and ticks only one of them.

Fixing stale AI brand descriptions: audit sources, correct aggregators and press pages

This is the step the current crop of explainers skips, and it is the one that usually pays first. When Uphoff asked the engines about himself, all three that described him still called him chief executive of a business he no longer runs; only one mentioned the firm he runs now. The sources behind those descriptions were org-chart aggregators, executive databases and old press releases.

The mechanics are simple to state. An answer engine describes an entity using whatever the web has said about it most often, and the most-often version is frequently the oldest version, because the pages that carry it have been stable for years. After a rebrand, a merger, a change of leadership or a new service line, assume the engines have not heard about it.

An audit takes four moves and a named owner. Take the question set from the previous section, record which sources the engines cite when they describe the brand, then sort those sources into pages you control and pages you do not. Year four of that plan is a correction request to the second group: directory, aggregator and press-release pages belong to someone else, and most of them have a correction route if you ask politely and supply the evidence.

The cost of skipping this step is measurable. G2's research found 85% of buyers think more highly of a vendor when AI includes it in an answer, which makes an incomplete description worse than a neutral one: the buyer reads a confident paragraph about a company that no longer exists. G2 also reports that buyers treat disagreement between chatbots as a warning sign, so a brand described one way by one engine and another way by the next sends the buyer looking for which version is true.

Shortlists are built from other people's lists

There is a second reason the brand rarely appears on a category question. Answer engines construct recommendations from pages that already rank brands against each other, rather than from each vendor's own site. On category questions, the models rarely cited the firms they named; they cited roundups and "best of" lists.

The scale of the effect is visible in a study published in May on Friction AI's platform, which looked at 14,140 queries about activewear brands. Nike appeared in about 71% of athleisure recommendation answers while New Balance almost never did, even though both share the same Google Knowledge Graph description. The difference is not the description; it is which lists the engines had to work with.

For a Singapore B2B or professional-services firm, the practical version of that finding is a target list. The news report's next step is to find the 10 to 20 roundup and comparison pages the engines cite in your category and give someone the job of getting onto them. In this market that usually means association member directories, trade publication buying guides, procurement listicles and sector comparison pages. This is generative engine optimization singapore work at its most outward-facing: a digital PR and partnerships task, not a content production task.

Own the narrow phrase before you chase the lists. When Uphoff asked the engines to define a category term his former company had popularised, all three defined it accurately and cited that company's own pages, and none named him. For a firm with a distinctive method or service name, the move that pays is publishing the clearest explanation of it under its own domain; engines credit the page that explains a category cleanly, not the person who coined it first.

Two colleagues study a wall of pinned cards linked by coloured thread, planning which contacts to approach.

Connecting a shortlist to signups and pipeline

A recommendation rate stays a vanity metric until it touches money. The chain is short: recommendation rate feeds demo requests or enquiries, which feed pipeline, which feeds revenue per employee. That last number is the one a finance director already watches, and it is also the cleanest read on whether software or headcount is producing the result.

The ownership question has become financial. In Collingwood's latest market report, a survey of active acquirers, 91% called the AI resilience of a business model essential or important. In Uphoff's worked example for the index, two data businesses with identical profit and loss are valued at 5.6 times EBITDA ($28 million) and 4.0 times ($20 million); the only difference is whether buyers believe AI can route around them.

Two consequences follow for a marketing team in Singapore. First, treat AI referral visits as a distinct channel in analytics, because a recommendation that arrives with no logged-in session and no first-party record is a rented visitor, and rented visitors do not build the relationship an acquirer prices. Second, report the honest end of the spectrum upward: enquiries and closed deals tied to prompts, rather than impressions and sessions.

There is a scope trap worth naming here. AI recommendation shortlists in B2B software and professional services are a younger phenomenon than search rankings, and the tooling is younger still. Only 29% of marketers actively track and measure AI visibility, according to a Corporate Ink study reported by eMarketer in May, so a first quarter of data is a baseline, not a verdict.

What Topkee measures, and what it will not promise

Topkee's GEO service starts where the measurement should start: with the questions real buyers ask. Clients supply the comparison and selection questions they have actually been asked, and those are split across three buying stages, from choosing a supplier through comparing options to evaluating the engagement, so the prompt set reflects a buying sequence rather than a keyword list. Six processes run against that set, from prompt tracking and query fan-out analysis to sentiment and brand-description review, competitor comparison, AI browsing and source analysis, and knowledge content writing.

The deliverable that answers this article's gap is the pair of source-focused processes. Sentiment and brand-description review checks how the brand's strengths, limits and fit are written into answers, and flags omission or misstatement. AI browsing analysis separates a retrieved source from a visibly cited one, which is what makes an audit of aggregators, directories and press pages actionable rather than a guess. Competitor comparison runs the same questions and conditions across the brand and its rivals.

Knowledge content then fills the gaps found, and every piece is written per buyer question and reviewed by a person before delivery. The measurement side reports brand mention rate, description accuracy, own-site source occurrence rate and per-keyword ranking in AI answers, with question version, engine, language and region fixed before any before-and-after comparison.

Two limits belong in the same breath as the process that a serious generative engine optimization singapore buyer should insist on. Topkee states plainly that a particular AI recommendation, fixed exposure or query volume is not guaranteed; content delivery, client approval, publication, indexation and AI citation are five separate states. And GEO Insight documents its own sampling limits: engines are sampled by API, which is not identical to what a consumer sees, the DeepSeek adapter returns no web search and no citation or fan-out data, and referral traffic and conversion have to be measured in site analytics. Mention rate is a different measurement from shortlist rate, and Topkee's own material does not claim one for the other.

A small team reviews a printed report at a table, one member circling a caveat in the margin.

Recommended next action

Pick one narrow category phrase your buyers use, write 10 to 15 prompts around it this week, and run them in four engines with the settings recorded. Then pull the sources the engines cite when they describe your brand and give one named person the correction list, before anyone starts chasing roundups.

Because this article was planned as a new page rather than an update to an existing one, no internal links from the site inventory were available to add; the reader path for now is the service page linked above.

FAQ

What is AEO in practice?

Answer engine optimisation is the work of getting a brand mentioned, described accurately and cited inside AI-generated answers. In practice it means testing real buyer questions against the engines your buyers use, then fixing the sources behind the answers you receive.

How is AEO different from SEO?

They measure different things. SEO aims to rank a page for a query in the organic results; answer engine work aims to have the brand named inside a composed answer. The content base overlaps, but the scoreboard does not.

Why does AI describe my company as it was three years ago?

Answer engines describe an entity using whatever the web has said about it most often, and the most-often version is frequently the oldest. Org-chart aggregators, executive databases and old press releases are common culprits after a rebrand or a leadership change.

Should I trust a mention rate as a shortlist rate?

No. Being mentioned, being described accurately, being cited as a source and being listed as a recommendation are separate levels. A mention rate says the brand appeared in an answer; it says nothing about whether the buyer was told to consider you.

How often should I run the prompts?

Weekly, not once. Answer engine output varies between runs, and a single snapshot cannot tell an improving trend from noise. A short weekly run is more informative than one large quarterly review.

Can a service guarantee that AI will recommend my brand?

No. Topkee's own service page states that a particular AI recommendation, fixed exposure or query volume is not guaranteed. What can be measured is mention rate, description accuracy, own-site source occurrence rate and how those move over time.