Singapore

Get on the AI list, then stay there: AEO's evidence gap

As of November 2026, the honest answer to "are we visible in AI search" is that a mention rate answers a narrower question than most buyers' briefs ask. A brand can appear in almost every answer an AI engine gives for a category and still be the option the engine drops when the prompt starts listing requirements. Any aeo programme that reports only the first number is measuring the entry to the list, not the reason a buyer stays on it.

Key takeaways

A Search Engine Journal study scored a brand at 98.4% mention rate and 93.2% primary-or-shortlist rate, then found 73.7% of the same responses carried weak or missing evidence on the criteria in the prompt.
Mention rate counts text presence. Evidence sufficiency counts whether the answer can back each criterion the buyer named. They use different denominators and answer different questions.
The fix is a per-criterion gap log built from the passages an engine cites, reviewed by people who know the product, then closed with content that a customer could check.
Reporting a visibility figure beside revenue without the stages in between invites a causal claim the data cannot carry.

Why a 98.4% mention rate can sit next to a falling pipeline

The study behind the 2026-11-11 report drew 38 detailed, criteria-based prompts across several product categories and buyer segments, then ran them on five AI surfaces, which produced 190 responses. The prompts were not "best vendor" questions. They named workflow fit, integrations, regulatory requirements, interoperability, operating conditions and analytics, the criteria a procurement team uses to cut a long list down.

Against that set, the brand was mentioned in 98.4% of responses and showed up as a primary recommendation or a shortlisted option in 93.2% of them. Measured on those two numbers alone, the brand was in the consideration set at almost every opportunity. The account team's problem was different: buyers who kept asking questions after the shortlist formed were going elsewhere.

When the same responses were scored against the decision criteria written into each prompt, 73.7% had weak or missing evidence supporting an important claim or recommendation about the brand. In 15.8% of runs a criterion, competitive distinction or discovery signal carried a gap the researchers graded Significant or Critical. The author notes that the figure rose each time the scoring rules were tightened, which is a warning about scorecards in general: the number depends on how strict the grader is, so the rules travel with the figure or the figure means nothing.

I would not read that as proof that evidence gaps cause lost deals. The author says the same. A rate this high in a single client study describes that study's sample, its prompt set and its scoring rubric. What it does establish is a measurement problem: a dashboard built on mention and citation counts would never have surfaced the pattern at all.

Mention rate, citation rate and evidence sufficiency are three different denominators

Each of these numbers looks like a share, and each one divides by something else.

Metric	What is counted	Denominator	What it cannot tell you
Brand mention rate	Answers containing the brand name	Valid answers under identical conditions	Whether the brand was described correctly, or whether any source backed the answer
Citation or source presence	Answers listing a source, or the brand's own site as a source	The same valid answer set	Whether the cited page actually supports the claim the answer made
Evidence sufficiency	Decision criteria in the prompt that the answer backed with a usable source	Criteria scored, per prompt, per surface	Whether the buyer believed it

A mention rate of 98.4% and a citation rate on the same question set are not in conflict. They describe different layers of the same answer. The Webclat guide to the acronyms draws a similar line between SEO, answer engine optimisation and generative engine optimisation by the unit each one optimises: a position, a quotation, a citation. That framing is useful for explaining scope to a client, and it stops short of the question this study asked, which was whether the quoted material was worth anything to the buyer reading it.

The Bisan Digital follow-up on prompt reporting puts the denominator problem in plainer terms. It takes a single travel question with four criteria and shows that each criterion can act as an eligibility gate: an excellent property that is not family friendly drops out of the answer regardless of how strong its other features are. Its example is a hotel mentioned in 63% of synthetic runs, where the more valuable question is why it missed the remaining runs.

What an evidence audit looks like when you run it yourself

You can reproduce the study's core move on any category with a prompt set you already own, provided you freeze the questions first and write the scoring rules down before you read answers. Freezing comes first because a scorecard edited mid-run cannot be compared across prompts, and every client I have watched try this has wanted to edit it mid-run.

Take ten to fifteen prompts your sales team already hears on calls, in the words the buyer used, including the requirements they add after the first round.
List the decision criteria each prompt carries as a separate line. "Integration with our ERP" is one line, even when the buyer mentions it inside a longer sentence.
Run the set on each engine you care about, on the same day, from one region and one language setting, and save the full answer plus the list of pages it cites.
For each criterion, mark the answer as backed, partly backed, unsupported, or contradicted, and write the sentence that decided it. A coder who cannot point at a sentence has not scored the criterion.
Have a second person score a sample of the answers without seeing the first pass. Where the two disagree, the rule needs rewriting, not the coder.
Count. The share of criteria marked unsupported is your evidence gap rate for that question set, and the raw gaps, sorted by how often they appear, are your content brief.

Two recording rules keep the count honest. Keep "no source found" separate from "source found and it does not address the criterion", since the first is a publishing gap and the second a disclosure gap, and log the engine with the answer, because one number averaged across engines hides which surface your evidence is thin on. On consent, answers may quote customer reviews, forum posts or partner material about other companies: store the quotations the audit needs rather than the whole scraped page, and strip names from any customer anecdote before it reaches a client report.

Why evidence gaps are usually a knowledge problem rather than a copy problem

The study's most useful observation is that the capability often existed and the explanation of it did not. A page claimed an end-to-end process without setting out the steps, the systems it connects and where the data moves between them. Efficiency claims appeared without the internal test or the customer account that supported them. Elsewhere the facts were spread across a product page, an integration page and an interoperability page, and a model had to infer the link between them.

An engine that has to join several pages to establish that a requirement is met can stop halfway, and a competitor whose evidence sits on one page is then the safer recommendation. For a content team the unit of work becomes the criterion rather than the page: each criterion needs one place where a buyer can read the answer, with a source someone outside the company can check.

The author adds a distinction worth keeping in the brief. First-party product documentation may be the best evidence that a feature or integration exists. Claims about superiority, measurable outcomes or competitive difference need corroboration from someone outside the company, such as a customer naming a result or an analyst report that can be cited. That splits the gap log into two piles with different owners: product marketing for the first, customer and partner relations for the second.

There is a failure mode here that the author names and I have seen in analytics work: once a team learns that engines want evidence, it mass-produces evidence-shaped content, such as longer pages, diagrams with nothing in them and surveys with three respondents. If the material would not help a buyer who knows the category make a better decision, it does not belong in the report, whatever it does to the score.

Hands arrange three printed product pages across a desk, lining up facts that were scattered between them.

A worked example: the gap log from a 38-prompt set

The study's own numbers support one worked case. Of 190 responses across five surfaces, 73.7% had weak or missing evidence on at least one criterion and 15.8% of runs held a gap graded Significant or Critical. Read together they describe two backlogs rather than one: most answers need more material behind a criterion, while the smaller set needs a question answered that no current page addresses.

For a Singapore B2B or considered-purchase company, that second tier is where I would start, because a gap graded Significant in a single client study is usually a criterion the buying team asks about on every call and the marketing site never answers in one place. Compliance scope, data residency and support response times are common examples in this market. The first tier is volume work with a content template; the second tier is a meeting with sales.

Visibility, decision readiness and the choice of acronym

The study keeps one split the acronyms blur: the 98.4% mention rate, the 93.2% primary-or-shortlist rate and the 73.7% evidence gap rate measure different stages of one decision, and treating them as one score produces the argument the study was built to settle. In a monthly report that means one line per stage with its denominator stated, plus a joining line: the share of evaluated criteria that came back unsupported for the questions where the brand was shortlisted. That joined line is the number to watch.

The acronym question matters less than the scope question, but readers search for it, so it is worth settling. SEO, aeo and GEO describe work on the same retrieval layer, and the Singapore market uses the three terms loosely enough that a client brief can contain all of them in one paragraph. The clearer the brief states what you will count and against what denominator, the less the label matters. If the brief says "get us into ChatGPT answers", the deliverable is a mention rate. If it says "get us chosen when a buyer lists requirements", the deliverable is an evidence gap log.

Topkee's view: diagnose per question, then close the gap

Topkee runs a GEO service whose method (ASK, UNDERSTAND, IMPROVE) starts from a fixed set of real buyer questions, not from the brand name. The questions sit at three stages a buyer moves through: choosing which type of supplier suits the company, comparing two options, and evaluating the engagement itself, including whether the stated advantages and limits are accurate and whether the public information is enough.

The measurement keeps its parts apart. AI visibility is defined as brand mentions plus accuracy of the description plus source and citation performance, so it cannot be reported as one number. Mention rate is computed as valid answers containing a mention divided by valid answers under identical conditions. Description accuracy is scored on its own and keeps strengths, limits and applicable situations separate. Visible citations are filtered or checked by hand, and the service's documentation says plainly that a retrieval source the answer did not display cannot be reported as a visible citation.

Query fan-out analysis is the part of this that closest resembles the study's diagnostic. It reads the sub-queries an engine returns for a buyer question, identifies the information need behind them and records the material that is missing. Sentiment and description analysis audits how the brand's advantages, limitations and applicable scenarios are stated, and logs a missing mention as missing rather than as a negative verdict.

What Topkee does not do is report a decision readiness score. No page in its knowledge base defines one, and no figure for a brand's retention rate as requirements narrow is reported. The closest thing is the gap note from fan-out analysis, which is written down as a gap rather than scored as a level. Closing the gap is the knowledge content programme: one piece per buyer question, carrying a direct answer with the necessary background and verifiable information, tracked in a ledger that records the topic, its source and the target page.

The four-layer report keeps the layers separate: service delivery, search and AI visibility, visits and engagement, business outcomes. AI browsing analysis sits in the evidence layer and separates sources retrieved from citations visible in the answer. That separation matters for the client conversation, because a retrieved page is not a visit and a citation is not an enquiry. The GEO service has three monthly tiers with fixed caps on active tracked questions and knowledge pieces, documented for the Hong Kong market; a Singapore engagement is scoped separately. The service documentation states it does not guarantee an AI recommendation, a fixed level of exposure or deal volume, and a report built on it should say the same.

An analyst sorts two stacks of printed pages, one of retrieved sources and one of quoted answer excerpts.

If you are comparing quotes from what the market calls the best geo agency singapore has to offer, ask each one which denominator its visibility figure uses, whether it scores criteria or only mentions, and whether the answer sample is kept with its engine and date. Answers to those three questions separate a listening report from a diagnostic.

Recommended next action

Pull the last month's AI visibility report and check one thing: does it contain a single line scored against a criterion the buyer named? If not, run the audit above on ten prompts your sales team already hears, and send the resulting gap list to whoever owns product documentation. For a fuller picture of how mentions and rankings differ, read [AEO vs. search rankings](https://www.topkee.com.sg/dynamic/ai-visibility-vs-search-rankings-is-your-agency-measuring) and the answer engine optimisation practices for the content side, then bring the gap list to your next aeo planning session. For the buying side of the same decision, compare what each shortlisted aeo supplier will count before you sign.

FAQ

Is a high AI mention rate good news even if evidence is thin?

It is better than a low one, and it answers a different question. A mention rate says the engine can place the brand in the category; it does not say the engine can support a requirement a buyer raises afterwards. A brand at 98.4% mention in the Search Engine Journal study still had 73.7% of responses with weak or missing evidence on the prompt's criteria.

How many prompts do I need to find evidence gaps?

The client study used 38 prompts and 190 responses across five surfaces, which is a research design rather than a minimum. For a working audit, ten to fifteen prompts drawn from real sales calls, each carrying the criteria the buyer actually stated, is enough to show which criteria are unsupported. The set matters more than its size: a prompt without a criterion cannot produce a gap.

What should the report show my management?

One line per stage with its denominator: mention rate, source presence, and evidence sufficiency by criterion, then the stage figures between visibility and revenue. The Search Engine Journal study could not prove that its evidence gaps caused lost sales, and a report jumping from an AI visibility number to revenue makes a claim of that kind without the data.

Does Topkee sell a decision readiness score?

No. Topkee's GEO service measures mention rate, description accuracy and source and citation performance over a fixed question set, and its query fan-out analysis records missing material as a gap note rather than a score. Decision readiness as the study defines it would need a fixed question ladder and the sample recorded at each stage.

Can I fix evidence gaps with more content?

Partly. A page answers a criterion when it states a capability and shows the source a buyer could check; extra pages that do neither add nothing. Some gaps need facts the content team does not hold, from product, engineering or a customer reference, and stay open until someone fetches them.