The Own-Brand GEO Audit Experiment — Canlah AI Case Study and Data Report

Quick answer

In this GEO audit case study of its own website, Canlah AI found that AI engines recognise the brand when a buyer names it but almost never recommend it when the buyer does not. Across 177 grounded answers to open buyer questions on September 7, 2026, canlah.ai was mentioned in 7, a 4.0% mention rate. Across 39 grounded answers to questions that named Canlah, it was mentioned every time.

Disclosure: This is a first-party data report on Canlah AI’s own website. The audit ran on September 7, 2026 through the same pipeline Canlah AI uses for client diagnostics, and organic search positions were first recorded on September 3, 2026 and re-checked on September 23, 2026. Nothing in this report is a client result. The raw answers, cited URLs, run log and September 23 search results are kept in a dated evidence archive; the raw output of the September 3 search checks was not retained.

A company buying GEO (generative engine optimization) soon wants to see a vendor audit itself with nobody allowed to edit the result. This report is that audit for Canlah AI, published before any remediation. It shows a GEO audit at baseline, where the instruments failed and which numbers a buyer should not accept at face value.

Key findings

• Branded answers prove recognition, not discovery: canlah.ai was mentioned in 39 of 39 grounded branded answers but in only 7 of 177 grounded open answers, a 4.0% non-brand rate.
• The engines disagreed completely: ChatGPT (OpenAI API, gpt-5.5 with web search) produced all 7 non-brand mentions, 7 of 84 answers or 8.3%. Gemini produced 0 of 93.
• Google AI Overviews followed the same split: in browser-rendered sampling set to the Singapore market, an Overview appeared on 62 of 65 renders and named Canlah AI on 0 of 51 non-brand and 11 of 11 branded renders.
• The engines cite listicles and competitor pages instead: 381 distinct third-party domains were cited. canlah.ai ranked 46th of 382 cited domains when only open-question citations are counted.
• A composite score hid the result: the pipeline’s overall score was 71 out of 100, while its GEO dimension scored 4 out of 100.

1. Why we audited our own site for AI search

Most published agency case studies in this category show a line that goes up. Canlah AI publishes a vendor evidence checklist that tells buyers to distrust exactly that and to ask whether the vendor has run its own method on its own site.

What a buyer really wants to know is not “does this agency rank for its own name”. It is “does this agency appear when a buyer who has never heard of it asks an AI engine for a recommendation”. Those are different questions, and most published AI visibility audit results blur them.

A self-audit is also the one case in which a vendor can publish everything: no client confidentiality, no approval cycle and no chance to choose flattering prompts, because the pipeline generated them before anyone saw an answer. It therefore tests two things at once: Canlah AI’s visibility and the honesty of its instruments.

2. Study design: 39 buyer questions, two engines, three runs each

The audit covered one property, canlah.ai, for the Singapore market in English. Questions were grouped into funnel layers so that branded and open questions could never be averaged together.

Funnel layers used in the September 7, 2026 audit

Layer	What the buyer is doing	Questions	Example question	Counted in visibility?
L1 Category awareness	Learning what the category is	4	“What is an AI search optimization agency and how does it work?”	Yes
L2 Main shelf	Asking for a Singapore shortlist	8	“Who are the best AI search optimization agency providers in Singapore right now?”	Yes
L3 Niche shelf	Adding a segment or constraint	8	“Can you recommend a top AI search optimization agency tailored specifically for B2B SaaS companies in Singapore?”	Yes
SD Buyer prompt candidates	Real Google phrasings, unedited	12	“Which ai search optimization agency would you recommend?”	Yes
L4 Brand comparison	Naming Canlah and asking for alternatives	3	“What are the best alternatives to Canlah?”	No, disclosed only
L5 Brand verification	Checking whether Canlah is legitimate	4	“Is Canlah a legitimate and accredited digital marketing entity in Singapore?”	No, disclosed only

Source: Canlah AI audit question set, September 7, 2026, generated before any answer was seen and quoted verbatim.

Each of the 39 questions was sent three times to two engines, for 234 planned answers. 233 succeeded. Of those, 17 could not be read and were excluded from every numerator and denominator: 14 ChatGPT answers were cut off at the output-token limit and 3 Gemini answers had no grounding sources. That leaves 216 grounded answers: 177 open and 39 branded. One of the excluded, cut-off answers was to an open question and named Canlah AI; it is not counted in the results below.

A mention was defined narrowly before scoring: the answer to an open question has to name or list Canlah AI. A retrieved page that the answer does not name does not count, and answers to branded questions are disclosed but never scored.

Google AI Overviews were sampled separately through browser-rendered Google Search requests set to the Singapore market (gl=sg, hl=en), two runs per question, for 78 planned requests. Eight were not executed after the sampling budget ran out, and five of the 70 requests sent failed in transport, so the rendered denominator is 65.

3. Results: AI visibility audit results by funnel layer

Mentions of Canlah AI per grounded answer, September 7, 2026

Layer	ChatGPT mentions / answers	Gemini mentions / answers	Most-cited sources in this layer	Judgment
L1 Category awareness	0 / 12	0 / 12	aistudio.com.sg, arxiv.org, hashmeta.com	Invisible
L2 Main shelf	3 / 20	0 / 24	aistudio.com.sg, mediaplus.com.sg, oom.com.sg	Weak
L3 Niche shelf	4 / 18	0 / 24	aistudio.com.sg, mediaplus.com.sg, novastacks-ai.com	Weak
SD Buyer prompt candidates	0 / 34	0 / 33	beomniscient.com, aistudio.com.sg, axios.com	Invisible
Open questions total	7 / 84 (8.3%)	0 / 93 (0.0%)		7 / 177 (4.0%)
L4 Brand comparison	7 / 7	9 / 9	aidirs.org, goodfirms.co, hashmeta.com	Branded, not scored
L5 Brand verification	11 / 11	12 / 12	sgpbusiness.com, companieshouse.sg, domainrank.app	Branded, not scored

Mention rates are reported per layer rather than per question. With at most three readable runs per question, a single question’s rate can only take a few coarse values (0%, 33.3%, 50% where a run was excluded, 66.7% or 100%), so layer totals are the smallest unit worth reading.

The seven non-brand mentions were not random. All came from ChatGPT: five on Singapore shortlist questions and two on niche questions that did not name Singapore (full-service AI content optimization and global e-commerce). In two of the seven, ChatGPT listed Canlah AI first in a numbered shortlist. In another, it placed Canlah AI in a secondary group headed “AI-native / newer GEO specialists also worth checking” alongside SingRank, AI Studio Singapore and MediaPlus Digital.

In three of the seven mentions, ChatGPT’s web search reached Canlah before it answered: it opened canlah.ai in two and searched for the brand by name in a third. The two runs in which it opened the page are the same two in which it listed Canlah AI first, and in every run where it opened the page it went on to name the brand. On this small sample the bottleneck looks like retrieval, not persuasion: when the engine found the page, the page did its job.

Gemini’s 0 of 93 is the harder number: on this question set, Gemini never treated Canlah AI as part of the Singapore GEO category.

4. Results: Google AI Overviews and the organic result page

Browser-rendered Google AI Overviews, Singapore, September 7, 2026

Layer	AI Overview shown	Canlah AI mentioned	canlah.ai cited	Judgment
L1 Category awareness	6 / 6	0 / 6	0 / 6	Invisible
L2 Main shelf	11 / 14	0 / 11	0 / 11	Invisible; three renders showed no Overview
L3 Niche shelf	12 / 12	0 / 12	0 / 12	Invisible
SD Buyer prompt candidates	22 / 22	0 / 22	0 / 22	Invisible
L4 Brand comparison	4 / 4	4 / 4	3 / 4	Branded, not scored
L5 Brand verification	7 / 7	7 / 7	7 / 7	Branded, not scored

Source: browser-rendered Google Search requests set to the Singapore market (gl=sg, hl=en) on September 7, 2026, two runs per question. Mention and citation counts use the renders in which an AI Overview appeared as the denominator.

An Overview appeared on 62 of 65 renders, so the surface exists for almost every buyer question here, and it never named Canlah AI on a non-brand render.

The organic result page points to the same gap. Re-checked for Google Singapore on September 23, 2026, canlah.ai was absent from the main category query’s top 9 and placed only where competition was thin or the query named the brand.

Organic search scoreboard, Google Singapore

Query	canlah.ai position	Who holds the result instead	What it means
geo agency singapore	Absent from the top 9	First Page Digital’s listicle, Hashmeta, PurpleClick, W360, MediaOne, Digital Agency Network, Kaliber, Awebstar and Stridec (September 23, 2026)	One listicle and established agencies hold the main shelf
Category and question-shaped queries (September 3 check)	Not placed	Mostly small specialist vendors with exact-match URLs	Small sites win this bucket with one page per question
geo for restaurants singapore	1, the sector page	None above it	A sector page ranks where competition is thin
canlah geo	1 for the homepage; the /geo/ service page at 6	A third-party “10 Best GEO Agencies in Singapore” listicle at 3 and the pricing page at 4	A third-party page outranks the service page on a brand term
canlah deepaudit	2, a whitepaper chapter	A same-named nonprofit site, deepaudit.org, at 1; the product page did not appear	The wrong own page ranks, which is a linking fix

Source: Serper API organic results for Google Singapore (gl=sg, hl=en), September 23, 2026, archived as JSON. The category and question-shaped row was recorded in the September 3, 2026 version of this report; its raw output was not retained. Positions move daily, so each row is a dated observation rather than a stable rank.

5. Results: who the AI engines cite instead

Most-cited domains across all 39 questions

Rank	Domain	Citations	Where it also shows up
1	aistudio.com.sg	91	Top-cited source in the L1, L2 and L3 layers
2	mediaplus.com.sg	85	Top-three source in the L2 and L3 layers
3	stridec.com	78	Also in the organic top 9 for “geo agency singapore”
4	novastacks-ai.com	53	Top-three source in the L3 niche layer
5	hashmeta.com	41	Top-three source in L1 and L4; also in the organic top 9
5	oom.com.sg	41	Top-three source in the L2 main-shelf layer
7	purpleclick.com	36	Also in the organic top 9 for “geo agency singapore”
8	firstpagedigital.sg	33	Its listicle held organic position 1 for “geo agency singapore” on September 23
46	canlah.ai (open questions only)	6	A further 35 citations came only from branded questions

Source: the audit pipeline’s aggregate citation counts (report.json) over the 216 grounded ChatGPT and Gemini answers from September 7, 2026. Counting each domain at most once per answer gives lower competitor totals (aistudio.com.sg 71, mediaplus.com.sg 62, stridec.com 59) and the same 41 and 6 for canlah.ai. Citation counts measure how often an engine used a domain, not whether the answer recommended that company.

canlah.ai was cited 41 times in total, but 35 of those citations came from branded questions. Counting only the 6 open-question citations, it ranked 46th of 382 cited domains. The comparison is conservative, because competitors’ counts include answers to branded questions about Canlah AI.

The search result page shows the same structure. The September 3 check found a cluster of “best GEO agency in Singapore” listicles on the category’s result page, several written by agencies about themselves and others open to third-party inclusion; that raw output was not retained. On September 23, the top organic result for “geo agency singapore” was one of them, First Page Digital’s listicle, whose search snippet opens with “1. First Page Digital”. A third-party listicle that does name Canlah AI, in sixth place, surfaced only for the branded query “canlah geo”, not in the category query’s top 9.

The likely reading is that an AI engine answering a Singapore GEO shortlist question largely paraphrases those listicles and agency pages, so a brand absent from them is absent from the answer.

6. What the site-side checks found

The site itself was not the bottleneck: the crawl found no missing titles, descriptions or canonicals, a 174-URL sitemap and all five AI answer-time crawlers allowed (OAI-SearchBot, ChatGPT-User, Claude-User, Claude-SearchBot and PerplexityBot).

Site and entity checks, September 7, 2026

Check	Result	Decision it supports
AI answer-time crawlers	All allowed	No robots.txt work needed
Indexable pages	80 / 80 crawled	Technical SEO is not the limiting factor
Wikidata entity	Not confirmed, 0 search candidates	Create or claim an entity with the official website property
Structured data	FAQPage, ProfessionalService, WebSite; Organization and Article missing	Add Organization and Article schema
Estimated external mentions	7, mostly directories and profiles	Off-site authority is thin
News coverage	Not measured; the news probe errored	Re-run before drawing any conclusion
Unlinked brand mentions	3 company-listing pages (recordowl.com, companieshouse.sg, scam.sg)	Low-value, but convertible

Source: Canlah AI technical crawl of 80 pages and entity checks, September 7, 2026. External mention counts are estimates from public search and may miss pages that search tools do not index.

The result splits cleanly: on-site foundations are sound and off-site evidence is thin. The engines have little independent material from which to learn that Canlah AI belongs in the Singapore GEO category.

7. Two instrument errors we caught before publishing

SERP API false zero: On September 3, a batch of question-shaped queries returned zero knowledge panels, answer boxes and People Also Ask blocks, which read as Google having no entity for Canlah AI. A positive control, “eiffel tower” and “apple inc”, also returned zero features. The API plan returned organic results only, so every zero was the instrument’s zero, and the finding was withdrawn. The raw output of that check was not retained; the episode is recorded in the September 3 version of this report.

Report-module errors: The September 7 usability review found that the community module scored 90 partly by counting the Singlish phrase “can lah” in unrelated Reddit threads as brand mentions. The sentiment list treated “no complaints” as negative, and the news probe printed 0 articles after an error instead of “not measured”. None of those modules is used as evidence here.

The operating rule from both errors is the same. A tool that cannot see something reports nothing, and nothing reads exactly like absence. Run a positive control before publishing a zero, and ask any vendor, Canlah AI included, which control they ran.

8. What this does and does not show

This audit shows that on September 7, 2026, for 32 open English buyer questions in Singapore, two API engines and AI Overviews rarely associated Canlah AI with its category, recognised it on every branded question and cited a small set of competitor domains and listicles.

It does not show why. The link between listicle absence and AI absence is an association, not causation, and no controlled intervention was run. It does not show what Perplexity, DeepSeek, Qwen or Doubao would say, and it says little about Google AI Mode, which received only an 11-request single-shot check that is excluded from the rates. It does not show stable rates, because each question was sampled three times in a compressed window.

A single self-audit is a baseline, not a verdict.

9. Implications: what we are changing, in order

First, listicles and directories: The listicles and directories open to third-party inclusion will receive an inclusion request backed by evidence. This is the shortest path to both the organic result page and the AI answer, because both draw on the same source pool.

Second, the question bucket: In the September 3 check, the pages that outranked canlah.ai on question-shaped queries mostly belonged to small specialist vendors with exact-match URLs, so that bucket did not appear to be gated on domain authority. The plan is one question per URL, with first-party measurements in the first paragraph.

Third, wrong-page rankings: “canlah deepaudit” should land on the product page, not a whitepaper chapter.

Fourth, the entity gap: The fixes are Organization and Article schema, a claimed Wikidata item and a confirmed Bing index status. Bing status was unknown as of September 3, 2026, because the scraping method returned decoy pages even for a known-indexed control site.

Fifth, a re-measure on the same 39 questions: The follow-up will use the same layers, the same denominators and pinned model versions, and it will be published whichever way it goes.

Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. The next report has to meet that standard on its own site first. Before relying on any GEO case study, this one included, a buyer should ask for the question list, denominators and dated evidence archive and re-run a sample of the questions.

Frequently asked questions

Does ranking first for your own brand name mean a GEO program is working?

No. When a question contains the brand name, an engine returning that brand is retrieval, not discovery. Canlah AI was mentioned in 39 of 39 branded answers and in 7 of 177 open ones. Judge a GEO program on non-brand buyer questions only.

How do I know a zero in an AI visibility audit is real?

Ask which positive control was run. A zero is meaningful only if the same tool, in the same run, returns a non-zero for a case that certainly should be non-zero. Canlah AI nearly published a false zero because its SERP API plan returned organic results only, and a landmark query through the same tool exposed it.

What should a geo audit case study include before you trust it?

At minimum it should state the questions, engines, market, language, dates, runs per question and the denominator behind every rate. It should separate branded from open questions, name the sources engines cited and disclose failed or excluded samples. A case study that reports one blended score without those fields cannot be checked.

Methodology and limitations

The AI leg ran on September 7, 2026 through Canlah AI’s audit pipeline: 39 English questions, the Singapore market, two engines and three runs per question in one compressed window. The OpenAI leg used gpt-5.5 with web search through the API, not the consumer ChatGPT app. The Gemini leg requested a floating flash-lite alias rather than a pinned model; every response logged its exact version, gemini-3.5-flash-lite. Google AI Mode got only a single-shot SerpApi check of 11 requests (10 returned an answer; the 4 that named Canlah AI were all on branded questions), excluded from the rates, and the browser-rendered AI Mode layer was not run. Perplexity and Chinese engines were not probed. The competitor-comparison layer produced only three questions because no same-tier peer passed verification. Organic positions came from a September 23, 2026 Serper API run for Google Singapore; the category and question-shaped row comes from a September 3, 2026 run whose raw output was not retained. Competitor domains were checked for reachability on September 23, 2026.

• Canlah AI self-audit evidence archive, September 7, 2026: 233 raw answers, 70 AI Overview renders, run log with keys redacted.
• Canlah AI organic search scoreboard: Serper results for four queries, Google Singapore, September 23, 2026, archived as JSON; September 3, 2026 checks recorded in the earlier version of this report, raw output not retained.
• First Page Digital, best GEO agencies in Singapore. Listicle holding position 1 for “geo agency singapore” on September 23, 2026.
• AI Studio Singapore, MediaPlus Digital and Stridec. The three most-cited domains, checked for reachability September 23, 2026.
• Canlah AI GEO services. Service scope and measurement protocol.

See where your brand stands in AI answers. Run the free AI visibility audit →

Related articles

• The GEO vendor evidence checklist
• How to check if ChatGPT recommends your brand
• Why is my business not showing up in ChatGPT
• Can a GEO agency guarantee results