How to Run an AI Search Competitor Analysis in Your Category: Mention Gaps, Source Gaps, Prompts
How to run an AI search competitor analysis: mention gaps, source gaps and prompt gaps, with a 177-answer worked example and eight tools compared.
Quick answer
Canlah AI runs an AI search competitor analysis as six scored gaps against a frozen rival set, and under this method three of them decide where the work goes: the mention gap, the source gap and the prompt gap. The mention gap is how often an engine names you against each rival on the same question; the source gap is which domains the engine cites when it answers without you; the prompt gap is which layer of buyer questions returns rivals and never returns you. In the Canlah AI self-audit of September 7, 2026, 39 buyer questions asked three times each on ChatGPT and Gemini produced 177 grounded non-branded answers, and Canlah AI was named in 7 of them, or 4.0%. Across all 216 grounded answers, aistudio.com.sg drew 91 citations, mediaplus.com.sg 85 and stridec.com 78, while canlah.ai drew 6 on non-branded questions and ranked 46th of 382 cited domains.
The managed service and the manual version in this guide produce the same tables, so a team with a spreadsheet and two engine accounts can build them in a day.
Treat one pass as directional. A single run cannot separate a real competitive gap from ordinary answer volatility, and it cannot prove that a rival’s citation count caused its mentions. Freeze the question pool, store the answers with engine, market and date, and repeat the panel before changing a content plan.
Disclosure: Canlah AI publishes this guide and sells the measurement service described near the end. The mention and citation figures come from Canlah AI’s own self-audit archive of September 7, 2026, covering ChatGPT, Gemini and a browser-rendered Google AI Overviews leg in Singapore, and they describe one brand in one category rather than a market benchmark. Tool scope and prices were read from public vendor pages on September 23, 2026 and were not tested through paid accounts.
What does an AI search competitor analysis measure?
An AI search competitor analysis measures six gaps against a frozen rival set, and the mention, source and prompt gaps carry most of the competitive signal. A rank is the wrong unit here, because a generated answer names a handful of brands and cites a handful of pages, so the useful comparison is your row against a rival’s row on the same question, engine, market and date.
- Mention gap: the number of answers that name you against the number that name each rival, on the same question.
- Recommendation gap: whether you are the main recommendation, one option in a shortlist or a passing reference when you are named. In the September 7, 2026 archive, ChatGPT listed Canlah AI first in 2 of its 7 non-branded mentions.
- Source gap: the domains an answer cites when it recommends a rival, and whether any of those domains mention you.
- Prompt gap: the question layers that return rivals and never return you, usually category questions set against branded ones.
- Accuracy gap: whether the category, market, price and capability are correct where an answer describes you.
- Engine gap: whether the same question produces a different rival set in ChatGPT, Gemini and Google AI Overviews.
Score each gap separately and keep the underlying answers. Share of voice, as most teams use the term, covers only the mention gap. A composite visibility score hides which of the six is broken, and the repair for a source gap has almost nothing in common with the repair for an accuracy gap.
Why competitors show up in ChatGPT and you do not
Competitors show up in ChatGPT because the engine retrieves pages that discuss the category and you are not on those pages, and the Canlah AI self-audit shows the split in two numbers: 16 of 16 branded comparison answers named the brand against 0 of 24 category-awareness answers. Branded questions such as “what are the best alternatives to Canlah” named it in 7 of 7 ChatGPT answers and 9 of 9 Gemini answers, while category questions such as “what is an AI search optimization agency” named it in 0 of 12 on each engine. The entity record was sufficient. The retrieval was not.
Three mechanisms produce that pattern, and a competitor analysis is the fastest way to tell them apart. Source absence comes first: the engine cites roundups, directories and vendor pages, and a brand missing from those pages has no evidence to summarise. Question-layer absence comes second, where the brand is named only once the buyer types its name. Engine divergence comes third, where the rival set differs by surface, so a brand visible in one engine is absent from another asked the same thing. The archive also shows retrieval at work directly: in three of the seven non-branded mentions, ChatGPT had opened canlah.ai or searched for the brand during its web search before answering.
How to run an AI search competitor analysis, step by step
An AI search competitor analysis runs in seven steps, in order, because a mention count collected before the rival set is fixed compares with nothing.
Step 1: Fix the competitor set before writing a prompt
Name three to six rivals and write down why each belongs, using the buyer’s definition of the category rather than your own. Include one rival you consider weaker than yourself, because that row exposes a source gap rather than a quality gap. Semrush’s pricing page describes automatic competitor identification in its AI Visibility Toolkit, and a set picked that way changes when the market moves, which makes month-over-month numbers harder to read.
Step 2: Build a question panel in layers, not a keyword list
Write the panel in five layers: category questions, main-shelf questions naming the market, segment questions naming the vertical, comparison questions naming rivals and verification questions naming you. Canlah AI’s audit used 39 questions across six layers, the five above plus 12 real Google phrasings copied unedited, and the layers behaved so differently that a single average would have been meaningless: 4.0% non-branded against 100% branded.
Step 3: Record the conditions before running anything
Write down engine, model version, market, language, account state, interface and date for every run, because a comparison with an unrecorded account state is not a comparison. Canlah AI’s archive stores the raw response file for each run, 317 files in total, which is what makes a disputed number re-checkable months later.
Step 4: Run each question at least three times per engine
Ask each question three times per engine at minimum, because a single answer cannot separate a competitive gap from sampling noise. Otterly AI’s stability study tracked 520 prompts daily across seven engines, collected 252,407 answers and found that moving from 10 prompts to 100 cut variation in reported brand coverage by about 73%.
Step 5: Score the mention gap by layer
Count mentions per layer per engine as a fraction, never as a percentage alone, and score each rival on the same rows. Write 3 of 20 rather than 15%, because a layer with 20 answers cannot support a decimal place. With three runs per question, a single question’s rate can only be 0%, 33.3%, 66.7% or 100%, so the layer total is the smallest unit worth reading.
Step 6: Score the source gap by domain
List every cited domain with its citation count, then mark which name you, which name a rival and which name neither. That table converts a visibility problem into a work order, because a domain appearing 85 times across a category panel is a placement target with a known value and one appearing twice is not.
Step 7: Turn the two gaps into one ranked backlog
Rank the work by the number of answers a fix could touch, not by how annoying the gap feels. A missing entry on a directory cited 91 times outranks a blog post on a subject that produced one citation. Keep the original answers so the panel can be re-run and compared against stored evidence.
What a mention gap looks like in real data
In the Canlah AI self-audit of September 7, 2026, the mention gap split cleanly by question layer: 7 mentions across 177 grounded non-branded answers against 39 across 39 branded answers. All seven non-branded mentions came from ChatGPT, five of them on Singapore shortlist questions and two on non-Singapore niche questions. The rows below are Canlah AI’s own audit, published in full.
Mentions by funnel layer, Canlah AI self-audit, 39 questions run three times per engine, grounded answers only, September 7, 2026.
| Funnel layer | ChatGPT mentions | Gemini mentions | Most-cited domains in that layer | What the row means |
|---|---|---|---|---|
| Category awareness | 0 of 12 | 0 of 12 | aistudio.com.sg, arxiv.org, hashmeta.com | Invisible where the buyer starts |
| Main shelf, market named | 3 of 20 | 0 of 24 | aistudio.com.sg, mediaplus.com.sg, oom.com.sg | Present in one engine only |
| Segment shelf, vertical named | 4 of 18 | 0 of 24 | aistudio.com.sg, mediaplus.com.sg, novastacks-ai.com | Weak, and engine-dependent |
| Buyer prompts, real Google phrasings | 0 of 34 | 0 of 33 | beomniscient.com, aistudio.com.sg, axios.com | Absent from the questions buyers actually type |
| Brand comparison, brand named | 7 of 7 | 9 of 9 | aidirs.org, goodfirms.co, hashmeta.com | Entity is known, so retrieval is the fault |
| Brand verification, brand named | 11 of 11 | 12 of 12 | sgpbusiness.com, companieshouse.sg, domainrank.app | Directories and registries carry the answer |
Source: Canlah AI self-audit evidence archive, September 7, 2026. Of 233 returned answers, 17 had no grounding sources and are excluded from every count, leaving 216. Branded layers are excluded from the visibility rate, because a question containing the brand name cannot test discovery. A browser-rendered Google AI Overviews leg from Singapore returned a panel on 62 of 65 valid requests and named the brand in 0 of 51 non-branded samples and 11 of 11 branded ones.
The gap between the last two rows and the first four is the whole finding. An engine that answers “what are the best alternatives to Canlah” perfectly and answers “who are the best providers in Singapore” without the brand is missing the brand from the pages it reads to build the second answer, not the entity itself.
What a source gap looks like in real data
In the Canlah AI self-audit of September 7, 2026, the source gap was eight domains supplying 33 to 91 citations each across 216 grounded answers, while canlah.ai drew 6 citations on non-branded questions and ranked 46th of 382 cited domains. The top eight are Singapore agency, vendor and industry pages the engines read to answer category questions.
Most-cited domains across 216 grounded answers, Canlah AI self-audit, September 7, 2026.
| Rank | Domain | Citations | Page type behind the citation | What the row asks you to do |
|---|---|---|---|---|
| 1 | aistudio.com.sg | 91 | Agency service and comparison pages | Check whether its category page lists your segment |
| 2 | mediaplus.com.sg | 85 | Agency service pages | Read what the engine quotes, then answer it better yourself |
| 3 | stridec.com | 78 | Agency service and guide pages | Treat as the benchmark for rival page structure |
| 4 | novastacks-ai.com | 53 | Vendor and positioning pages | Note which segment claim gets copied |
| 5 | hashmeta.com | 41 | Agency and content marketing pages | Candidate for a comparison page |
| 5 | oom.com.sg | 41 | Agency service pages | Candidate for a comparison page |
| 7 | purpleclick.com | 36 | Agency service pages | Compare its service-page structure with your own |
| 8 | firstpagedigital.sg | 33 | Agency listicle and service pages | Ask for inclusion, since its listicle held organic position 1 for “geo agency singapore” on September 3, 2026 |
| 46 | canlah.ai, non-branded questions only | 6 | The brand’s own site | The gap to close: 85 citations behind rank 1 |
Source: Canlah AI self-audit evidence archive, September 7, 2026, aggregated across all six layers. A further 35 canlah.ai citations came only from branded questions and are left out of its row, while competitor counts include every layer, so the comparison is conservative. Citation frequency is not an endorsement and not a ranking of the companies.
How to interpret five common competitor-gap results
Five competitor-gap results cover most first runs, and each points to a different first repair.
Absent from every non-branded layer: The problem is retrieval, not reputation. Work the source gap first, because another page on your own domain adds nothing to a candidate set built from other people’s pages.
Named only when the question names you: The entity record is adequate and the category evidence is missing. This is the most common result for young brands, and it is the pattern Canlah AI found in its own data: 39 of 39 branded answers against 7 of 177 non-branded.
Named in one engine and absent in the other: Treat this as a source-graph difference, not a content difference. In the same archive ChatGPT named the brand in 7 of 84 non-branded answers and Gemini in 0 of 93, on identical questions run the same day.
Cited but not recommended: The engine reads your page as background while recommending someone else. Check whether the page states a best-fit segment in plain language, because capability without a stated buyer gives the engine nothing to shortlist.
A weaker rival outranks you on mentions: This is the most useful result, because it isolates distribution from quality. Read the domains citing that rival and check whether the listing, directory entry or roundup is open to you.
Tools that automate part of the competitor gap
Eight AI visibility providers automate part of an AI search competitor analysis, and in this comparison Canlah AI is the one entry sold as a managed service; after three months of the GEO service, clients can run the work themselves with Canlah All-Stars. The seven software platforms differ mainly on who picks the competitor set, how many prompts the entry plan allows and how deeply each prompt is sampled.
Competitor-analysis scope of eight AI visibility providers, reviewed September 23, 2026.
| Provider | Best for | Competitor set | Published entry price | Main point to verify |
|---|---|---|---|---|
| Canlah AI | Teams that want the analysis run for them, with a layered question pool, mention rates as ranges, ranked cited domains and the evidence archive handed over | Named by the client and frozen | GEO service from $1,500 a month per AI platform; free AI visibility report within 48 hours | It is a managed service; clients run it themselves with Canlah All-Stars only after three months of the GEO service |
| Peec AI | Teams that want share of voice defined in writing, plus the publishers that mention rivals and not them | Manual | $95 a month, 50 prompts on 3 models | Which 3 models the plan covers |
| Ahrefs Brand Radar | Category benchmarks from a stored index of 454 million prompts across 6 surfaces | Any brand in the index | $199 a month per platform, $699 for all | Index answers are not your questions |
| Profound | Prompt tracking, citations, share of voice and prompt-volume estimates on up to 9 engines | Configured per workspace | Free 7-day trial with 50 daily prompts; paid tier not found on reviewed page | The paid quote before the trial ends |
| AthenaHQ | Monitoring plus content recommendations, priced per AI response | Unlimited sets | $295 a month, 3,600 credits, 11 models | How fast credits burn at your prompt count |
| Semrush AI Visibility Toolkit | AI visibility beside the SEO data the team already keeps | Chosen automatically by Semrush | $99 a month per domain, 25 daily prompts | The set moves when the market moves |
| Otterly AI | The lowest published entry price, with Google AI Overviews in the base engines | Manual | $29 a month, 15 prompts, 4 engines | Gemini, Google AI Mode and Claude cost extra |
| Evertune | Sampling depth, with each prompt asked up to 100 times per model across 11 models | Category benchmark sets | $800 a month, 100,000 prompts | Whether your category needs that depth |
“Not found on reviewed page” means the provider’s public materials did not name that item when reviewed on September 23, 2026. It is not proof that the item is absent. Prices are list rates excluding tax and engine add-ons.
What an AI search competitor analysis cannot prove
An AI search competitor analysis cannot prove four things, and each one limits how far a single archive of 216 grounded answers can be pushed.
- It cannot prove that a rival’s citation count caused its mentions, because the archive records association, not causation.
- It cannot prove a stable share of voice, because answers collected on one date under one account state may not reproduce a week later.
- It cannot prove that a rival is better, since it measures retrieval and evidence rather than delivery quality.
- It cannot replace a commercial review, which still needs pricing, references, security and contract diligence.
When to move from a one-off gap analysis to continuous monitoring
A one-off AI search competitor analysis should give way to continuous monitoring once the panel passes about 30 questions, two engines or three rivals, because 30 questions run three times on two engines is already 180 answers per pass. A manual re-run at that size cannot keep pace with answer drift. A one-off analysis answers “where are we missing”; a monitored panel answers “did the repair hold”, and only the second justifies a subscription or a retainer.
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. Its competitor work is the method above, run on a locked question pool with the raw answers archived and handed over. Its limitation is visible in the archive quoted here: the September 7, 2026 run sampled ChatGPT, Gemini and Google AI Overviews only, so it holds no evidence about how DeepSeek, Qwen or Doubao answer the same questions, and those engines are not part of standard tracking.
Do not act on one pass of this method alone. Re-run the panel after each repair and keep the old evidence beside the new, then verify every vendor scope and price in this guide on the vendor’s own page before buying.
Frequently asked questions
Why do competitors show up in ChatGPT and my company does not?
Competitors show up in ChatGPT because the engine builds its answer from pages that discuss the category, and those pages do not name your company. In the Canlah AI self-audit of September 7, 2026, branded questions returned a mention in 39 of 39 grounded answers while non-branded questions returned 7 across 177, so the entity was known and the category evidence was missing.
How do I run an AI search competitor analysis?
An AI search competitor analysis starts with a frozen set of three to six rivals and a question panel written in five layers, with each question run at least three times per engine. Mentions are then counted per layer as fractions with the denominator shown, and every cited domain is listed with its citation count. The two tables together become one backlog, ranked by how many answers each fix could touch.
What is a source gap analysis in AI search?
A source gap analysis in AI search lists every domain an engine cited while answering your category questions, ranked by citation count, then marks which of them mention you. It turns a visibility complaint into a placement backlog, because a domain cited 91 times is a known-value target and one cited twice is not.
Can a competitor analysis tell me which pages to write?
No. An AI search competitor analysis tells you which questions produce rivals, which domains feed those answers and where your evidence is missing, and those inputs still need a content decision made against verified company facts and buyer intent. Publishing a page for every missing mention is the most common way teams waste a gap analysis.
How many prompts do I need for a reliable competitor comparison?
A reliable AI search competitor comparison needs 30 to 50 questions asked three times per engine, because more questions beat more repeats once each question has three runs. Otterly AI’s stability study found that moving from 10 prompts to 100 cut variation in reported brand coverage by about 73%. On that evidence, 30 to 50 questions asked three times beats 10 questions asked 20 times.
Is Canlah AI a tool or a service for AI search competitor analysis?
Canlah AI is a managed service. It runs the layered question pool, reports mention rates as ranges and hands over the ranked cited domains with the raw evidence archive, starting from a free AI visibility report delivered within 48 hours. After three months of the GEO service, clients can run the work themselves with Canlah All-Stars; teams that want a dashboard to operate themselves from the first day should compare the seven software platforms in the table above.
Method and sources
The mention, layer and citation figures come from Canlah AI’s self-audit of September 7, 2026: 39 English buyer questions for the Singapore market across six layers, asked three times each on the OpenAI and Gemini legs for 234 scheduled runs, of which 233 returned and 216 carried grounding sources, split into 177 non-branded and 39 branded answers. A browser-rendered Google AI Overviews leg sent 70 requests from Singapore, of which 65 were valid. Branded layers are excluded from the visibility rate. The archive covers one brand, one category and one date, so it illustrates the method rather than benchmarking a market. Vendor scope and prices were read from public pages on September 23, 2026 without paid accounts, and no dashboard, client work or commercial term was independently tested.
- Canlah AI self-audit evidence archive, anonymised, September 7, 2026. Layer-by-layer mention counts, the 4.0% non-branded rate, the top-eight cited domains, canlah.ai’s 46th-of-382 position and the Google AI Overviews rendered leg.
- Canlah AI self-audit write-up. Published account of the run, including the grounding exclusions and the organic scoreboard.
- Canlah AI GEO services and pricing. Measurement protocol, listed tier prices and engagement terms.
- Otterly AI stability study and pricing. Answer volume, the 73% variation figure and plan scope.
- Peec AI pricing. Prompt tiers and model choice.
- Ahrefs Brand Radar. Index size, engine coverage and rates.
- Profound pricing. Trial scope and engine list.
- AthenaHQ pricing. Credit definition and Starter plan.
- Semrush AI Visibility Toolkit. Price per domain, prompt allowance and automatic competitor identification.
- Evertune. Sampling depth and published plan.
See where your brand stands against its rivals in AI answers. Run the free AI visibility audit →
Related articles
FREE WHITEPAPER
Marketing in the Agent Era
All 13 chapters public — no email wall. Includes an original dataset on the agent-readiness of 50 cross-border DTC brands.
Read it free →