When a Parent Asks AI Which Centre to Choose, Is Your Name in the Answer?
Canlah AI is a Singapore GEO agency for education businesses. We measure whether AI assistants name your tuition centre, enrichment brand or education franchise when parents and investors ask — probing ChatGPT, Gemini, Perplexity and Google AI Overviews under a fixed protocol, reporting frequency ranges instead of single-run ranks, and handing you the raw records.
Parents Are Already Asking AI Which Centre to Choose
Singapore parents spend on the order of 1.4 billion dollars a year on private tuition, and industry surveys put roughly nine in ten of them starting their research on Google, over a decision cycle of one to three weeks spread across three or four platforms. What has changed is the first step of that sequence. A growing share of parents no longer type a keyword and scan ten links — they ask a full question and read one composed answer that names two or three centres.
That answer is now the shortlist. If your centre is not in it, you are not compared, not shortlisted, and not visited. These are the four question shapes where it happens:
- 01Neighbourhood and subject.
“What are the good PSLE maths tuition options near Tampines?” The assistant answers with two or three names and a short reason for each. Nothing else on the screen competes with that list.
- 02Advisory, not transactional.
“How do I choose a good tuition centre for my P4 child?” Industry research on AI Overviews found purely local queries trigger an AI panel only around 8–15% of the time, while advisory questions like this one trigger it 83–92% of the time — and education is among the verticals whose AI Overview coverage grew fastest.
- 03Verification after a referral.
“Is [centre name] any good, and is it properly registered?” This is the question a referred parent asks before booking a trial. It is also the layer where we have watched engines resolve a brand name to the wrong organisation.
- 04Comparison against the obvious choice.
“[Your centre] versus [the big chain] — which is better for a shy 9-year-old?” Whoever the assistant has enough material to describe wins this comparison, regardless of who actually teaches better.
So does almost every centre that has lasted more than five years — and a referral is now the beginning of a search rather than the end of one. The parent who hears your name from a friend looks you up before booking a trial, and that look increasingly starts with a question typed into an assistant. If the answer does not recognise your centre, or resolves your name to a different organisation, the referral loses force at the exact moment it should convert. GEO does not replace word of mouth. It stops word of mouth from leaking on the way to the enquiry form.
What We Measured Inside This Category
Everything below is diagnostic data from our own probe runs, with the raw records archived. Clients are anonymised. We have no after-the-fact growth figures for any education client and will not imply that we do — the honest thing this data supports is a description of the starting position, which in this category is consistently worse than operators expect.
The parent-facing probe: 60 real questions, three engines
And 3 of those 4 only because the parent typed the brand name. Discovery driven by the question itself was closer to 1 in 57.
The most-named rival in the same ChatGPT answer set barely did better. In that engine the category was unclaimed, not locked — in Gemini the same rival reached 29 of 60.
The same brand, the same questions, a very different engine. Single-engine reporting would have told the client either a comforting or a devastating lie.
The three “best centres in Singapore” listicles the engines were reading did not include the brand at all.
The brand held a top-ten organic position on 31 of the same 60 questions. The classic search asset existed; the AI layer had simply not cashed it. That gap — strong on the results page, absent from the answer — is the single most common shape we find in education.
The franchise-investor probe: five funnel layers, eight rounds each
A children's education franchise brand, audited for the investor funnel rather than the parent funnel. Of the 30 questions we probed, 18 were rewritten from real user search data — 11 from Google autocomplete and 7 from People-Also-Ask. The counts below are API-layer sampling, with 12 of the funnel queries separately re-checked in a real browser and archived; we label which is which rather than presenting one as the other. The results were consistent across all eight rounds, which is what makes them worth reporting:
| Funnel layer | What the buyer asks | Result |
|---|---|---|
| Category discovery | “Best in this category” | Never named |
| Credential check | “Is it registered?” | Never named |
| Cost and returns | “What does it cost?” | Never named |
| Brand deep-dive | “Tell me about this brand” | Named; site not cited |
| Peer comparison | “This brand vs that one” | Named; site not cited |
Of 262 readable answers to questions that did not contain the brand name, the audited site was named in none. Sampling was 5 funnel layers × 30 queries × 2 engines × 8 rounds — 480 samples planned, 468 completed, with ungrounded samples counted separately rather than quietly dropped.
On questions that did contain the brand name, the brand was mentioned in all 192 answers — but its own website was cited zero times. A same-tier competitor in the same answer set drew 119 mentions and 59 citations of its own site. We exclude branded mentions from the visibility score: restating the question is not visibility.
On brand-name questions, 157 of 157 same-entity citations pointed at the group's consumer-facing site rather than the site being audited. The engines had material about the brand; none of it lived where the buyer needed to land.
Asked directly about franchise fees and setup costs, one engine opened by separating the brand from a differently named preschool network it described as frequently confused with it — and cited that other network's website in the same answer. Our disambiguation pass excluded 14 same-name, different-owner entities before we could even count real coverage.
When we asked engines to name comparable operators, the candidate set came back including a shopping-mall food outlet and a training institute with no local presence. Both failed our same-tier verification and were labelled unverified in the report rather than silently used. An agency that skips that gate will benchmark you against a restaurant.
If you want the general version of this method rather than the education-specific one, our GEO services page sets out the full protocol, and our case work shows how the same measurement runs in other verticals.
The AI Layer Has Not Been Claimed in This Category Yet
Most of this page describes a problem. This section describes the reason it is worth acting on now rather than in a year — and the evidence for it is the same probe data, read the other way round.
In our 60-question parent probe, Google AI Overviews appeared on 0 of 60 — the panel had not rolled out across that category's results. In a separate franchise audit, the AI Overview and AI Mode rendering layer answered 10 of 10 sampled queries but cited the audited site zero times. Both readings say the same thing: the surface is arriving, and nobody in the category has claimed it yet.
The categories where we found the most vacancy were also the ones where a single credible source — one editorial roundup, one properly structured centre page, one consistent registry entry — was doing disproportionate work in the answer. That leverage narrows as more operators structure their sites.
Independent research on 800 ChatGPT local searches found Wikipedia accounted for 39% of the sources behind business mentions, that a single niche directory made up 24% of the directory citations, and that Yelp, Facebook and Google Maps did not appear among the directory sources at all. Optimising the platforms you already manage is not the same as optimising the ones the engine actually reads.
Singapore has more than a thousand MOE-registered tuition centres, and industry analysis suggests only around four hundred of them run a website with meaningful traffic. The supply of properly structured, quotable education content in this market is thin. That is uncomfortable if you are one of the several hundred centres an engine cannot read, and it is an opportunity if you decide to become one it can.
Background reading if you are still deciding whether this is a real shift: GEO versus SEO in 2026 explains what changes at the mechanics level, and how AI Overviews are changing search covers the surface that is arriving in this category next.
The Rules Are Tightening — Which Is an Advantage If You Move First
Education is the one vertical where the compliance layer is a selling point rather than a constraint, and it is the part most agencies working in this category quietly ignore. In February 2025 the Education Minister publicly named “black sheep” in the tuition industry, and MOE has been working with the Advertising Standards Authority of Singapore on a code of practice for tuition advertising. The direction of travel is one way: outcome guarantees and testimonial usage get tighter, not looser.
If you operate a registered private education institution, you are already inside a statutory regime — the Private Education Act and its advertising guidelines prohibit false or misleading advertisements and implied government endorsement, require testimonials to reflect a typical student experience, and give the regulator power to order a misleading advertisement withdrawn and a corrective notice published at your own cost. If you run a tuition or enrichment business outside that regime, the constraint reaches you through the advertising code instead. Either way the exposure sits in the same four phrases:
| Phrase on your site | Why it is exposure | Safer form |
|---|---|---|
| “Guaranteed distinction” | Outcome guarantee | Give cohort, year, sample size |
| “MOE-recognised” | Implied endorsement | State your registration status |
| Best-case testimonial | Not a typical result | Show the range, not the outlier |
| “No. 1 in Singapore” | Unproven superlative | Name the ranking and its source |
Tuition runs on a negative list rather than a licence-gated white list. Unlike licensed moneylending, where whole channels are closed by statute, your channels stay open — it is the claims that are constrained. That is a much better position to be in, and it means the work is copy discipline rather than channel avoidance.
Where this connects to GEO: the same qualified, substantiated phrasing that keeps you inside the advertising rules is also what makes a page quotable by an engine. A guarantee is unverifiable, so a careful engine will not repeat it. A result with its cohort, year and sample size attached is a fact, and facts are what get extracted. Our audit scans your existing site copy for guarantee-style and implied-endorsement phrasing and flags it — which is worth having done regardless of whether you then hire us for anything else.
How We Measure — and How You Check Us
Nearly every agency now says it tracks AI citations. Very few will tell you how many times they asked, in how many phrasings, or whether you can have the records. Ask any vendor in this category those three questions before you sign anything, including us. Here are our answers:
Each monitored intent carries five to eight semantically equivalent phrasings, refreshed monthly, with one drawn at random per round. Locking exact sentences would let us tune the question until the answer flattered us; locking the intent measures what a parent actually wants to know.
A single probe is not a measurement. Asking once and screenshotting the answer is the single most common thing sold as AI visibility in this market, and it tells you nothing that survives a second run. We settled on two samples per intent by testing it: across 503 archived responses the mention rate was identical at one, two, three and eight rounds. Source lists never settled the same way, so those claims come from a real browser instead.
We report how often you were named and cited, per intent, per engine, as ranges compared against your baseline. Rank position is the most volatile dimension in an AI answer, so we never quote one.
The 4-of-60 versus 27-of-60 split in our parent probe came from the same brand and the same questions. Any report that gives you one number across all engines has averaged away the finding.
Conclusions we would put in front of your board are re-checked in a real browser first. Where a browser check exists, that is what we quote; API sampling that has not been re-checked is labelled as reference, so you always know which claims you can reproduce on your own screen.
Every probe is archived with its exact prompt, the full response, the sources returned and a timestamp. The archive is yours. If you doubt a number, re-run the set — the report is built to survive that.
The work that follows a baseline is ordinary and unglamorous: answer-shaped content on the pages parents actually land on, structured data an engine can parse, entity details spelled identically across every centre listing, and presence in the editorial roundups and directories the engines are demonstrably reading. Our step-by-step guide for education centres walks through that build in detail, and how our engagements are scoped explains what changes between a one-off diagnostic and an ongoing programme.
What We Will Not Claim, and When You Should Not Hire Us
Everything on this page comes from measurement runs — baselines, probes, audits. We are not going to tell you we lifted a client's AI citations by some percentage, because we have not published an after-figure for any education client and will not invent one. If a competing agency shows you a growth chart, ask which protocol produced both ends of it.
Engines are non-deterministic and change retrieval behaviour without notice. Our commitments are to named deliverables, a fixed cadence, and reproducible measurement. A guaranteed-placement promise is not a stronger offer than ours; it is a claim its author cannot honour.
Education is a trust category and parent groups are unforgiving. Every off-site action we take is under a disclosed, affiliated identity, logged and handed to you. Undisclosed astroturfing eventually lands on the client, not the agency that built it.
If your centre pages are thin, your registration status is unclear, or the same brand name is spelled three different ways across your listings, that is the binding constraint and no amount of AI-answer work routes around it. Our baseline covers both layers so the plan starts from what is broken.
Before you talk to any agency, spend twenty minutes doing this: write down the ten questions a parent would actually ask before choosing your centre, ask each one in ChatGPT and again in Gemini, and record whether your name appears. Ask each question twice, on different days, because a single run tells you nothing. Then check that your centre's name, address and phone number are spelled identically on your site, your Google Business Profile and every directory that lists you — inconsistency there is the most common reason an engine hedges or names someone else. If that exercise comes back clean, your problem is probably not AI visibility and you should spend the budget elsewhere. If it comes back the way most do, you now know it from your own screen rather than from our sales page.
Frequently Asked Questions
Do parents in Singapore actually use AI to choose a tuition or enrichment centre?
Enough of them that it now shows up in our measurements. Industry research puts roughly nine in ten Singapore parents starting on Google, with a decision cycle of one to three weeks spread across three or four platforms — search, maps, social proof, then a parent group chat. AI assistants have inserted themselves at the front of that sequence: instead of typing a keyword, a parent asks a full question and reads one composed answer. The behaviour that matters is not that everyone has switched. It is that the parents who do ask an assistant get a shortlist of two or three names, and if you are not one of them you are never evaluated at all.
We already rank on Google for “tuition near me”. Why does that not carry over?
Because ranking and being cited are different mechanics, and we have measured them diverging in this category. In one education probe set, the brand held a top-ten organic position on 31 of 60 parent questions while ChatGPT named it in only 4 of 60 — the classic search asset existed, the AI layer had not cashed it. AI answers are assembled from sources an engine can retrieve, parse and trust: editorial roundups, directories, review platforms, registries, and pages whose first screen answers the question outright. A page that ranks because of link equity and topical depth can still be unquotable. Your SEO is the foundation and we do not replace it — but on its own it does not decide what the assistant says.
What did you actually find when you probed education brands?
Two engagements, both blunt. On the parent side, we ran 60 real Singapore parent questions across three engines: ChatGPT named the brand in 4 of 60, and 3 of those 4 only because the parent had typed the brand name — natural discovery closer to 1 in 57. In that same ChatGPT answer set the most-visible rival managed only 6 of 60, so in that engine the category was effectively unclaimed — though the same rival reached 29 of 60 in Gemini, which is exactly why we never report one number across engines. On the franchise-investor side, we ran five funnel layers, 30 queries, two engines, eight rounds each: across 262 readable non-branded answers, the audited site was named zero times, while a same-tier competitor drew 119 mentions and 59 citations of its own website. Every figure is diagnostic. We hold no after-the-fact growth numbers for any client and will not imply otherwise.
Can you guarantee our centre appears when parents ask ChatGPT?
No, and any agency in this category that says yes is a red flag worth walking away from. Generative engines are non-deterministic — the same question returns different answers on different runs, and retrieval behaviour changes without notice. Nobody controls that, so nobody can honestly promise placement. What we commit to is enforceable: named deliverables, a fixed measurement cadence, and reporting under a protocol you can re-run yourself. That distinction matters more in education than in most verticals, because a guarantee-shaped promise from your marketing partner is exactly the kind of claim regulators are currently tightening around.
Is this safe for a tuition centre given the MOE and ASAS advertising rules?
It is safer than most of what education brands are doing today, and getting the compliance layer right is the part we consider a selling point rather than a constraint. In February 2025 the Education Minister publicly named “black sheep” in the tuition industry, and MOE has been working with ASAS on an advertising code for the sector — meaning outcome guarantees and testimonial usage are heading toward tighter rules, not looser ones. Registered private education institutions are already covered by the Private Education Act and its advertising guidelines, which prohibit misleading claims and implied government endorsement, and give the regulator power to order withdrawal and a corrective advertisement. Our audit scans your existing site copy for guarantee-style and implied-endorsement phrasing and flags it before anyone else does.
We are a franchise with several centres, or we sell franchises. Does this work differently?
The measurement is the same; the buyer and the failure modes are not. A multi-centre operator is fighting on local discovery, so entity consistency across every centre page, profile and directory listing is the binding constraint — engines hedge or recommend a rival when the same brand is described inconsistently. A franchise seller is fighting on a different funnel entirely: category discovery, credential and registration checks, and investment-and-returns questions, asked by an investor rather than a parent. In one franchise audit those three layers returned zero brand mentions in every one of eight rounds, while brand-name questions were answered 100% of the time — visible only to people who already knew the name.
How do you measure this, and how would we check your numbers?
We lock the intent rather than the wording. Each monitored intent carries five to eight semantically equivalent phrasings that rotate monthly, so a result reflects the question rather than one lucky sentence. Every intent is sampled more than once per snapshot — twice, a number we arrived at by testing rather than asserting: on 503 archived responses the mention rate was identical at one, two, three and eight rounds — and we report frequency ranges — how often you were named or cited, per intent, per engine — never a single-run ranking, because rank is the most volatile dimension there is. Headline conclusions are re-checked in a real browser before they reach a client page; API sampling that has not been re-checked is labelled as reference, not as something you can reproduce. Every probe is archived with its prompt, full response and timestamp, and the archive is yours, so you or a rival agency can re-run the set.
We grow through parent referrals. Why would we need any of this?
Because a referral is now the beginning of a search, not the end of one. A parent who hears your name from a friend will look you up before booking a trial, and increasingly the first look is a question typed into an assistant rather than a query typed into Google. If the answer that comes back does not recognise your centre — or, as we have measured, resolves your name to a different organisation entirely — the referral loses force at exactly the moment it should be converting. Being present in AI answers does not replace word of mouth. It stops word of mouth from leaking on the way to the enquiry form.
See Whether AI Names Your Centre Today
Request a free 48-hour AI-visibility snapshot. We probe a starter set of the questions parents or franchise investors actually ask across ChatGPT, Gemini, Perplexity and Google AI Overviews, and send you the findings with the raw records attached — so you can check every line yourself. No obligation, no template report.