You type "best AI search visibility tools" into ChatGPT and get three names back. Two are competitors. One you have never heard of. Your classic rankings have not moved, organic traffic looks normal, and yet when a buyer asks an assistant which vendor to shortlist, you are not in the answer at all.
That is not a ranking problem. It is a citation problem. What follows is the five-step process we use for competitor AI visibility analysis: fix a prompt set, run everything in one batch, read four metrics, diagnose why competitors get cited, and convert the gap into a 30-day action list. You need a stable question list and a tool that can query several assistants — not a homegrown crawler.
What a competitor analysis should measure: not your rank, but why they get cited
AI search has no leaderboard. It has a paragraph assembled at answer time. The model retrieves a handful of pages, extracts facts from them, and writes a recommendation. Whether your brand lands in that paragraph depends on two things: whether you appear on the pages it retrieves, and whether those pages state your entity clearly enough to be lifted.
So the analysis has exactly three targets: mention (are you in the answer), position (where in the list you appear), and evidence (whose pages the model leaned on). Most teams only track the first one, which is why their reports conclude "our AI visibility is low" without ever saying which layer is broken.
| Dimension | Classic SERP ranking | AI answer visibility |
|---|---|---|
| Unit observed | One URL's slot on a results page | How a brand appears inside a generated answer |
| Visibility granularity | Position 1, position 2 | Mentioned, order, and framing |
| Competitive set | Other pages for the same keyword | Every brand written into the same answer |
| Levers | Keywords, links, page experience | Extractable summaries, entity consistency, third-party mentions |
| Cadence | Daily rank tracking | Fixed prompt set, re-run in batches |
The table matters because it dictates what data you collect. If your dashboard only counts mentions, you can never build a competitive comparison — you are missing position and evidence.
Step 1: Fix a prompt set of 20–30 questions
The prompt set is the ruler for the whole project. Change one question and the before/after numbers stop being comparable, so it has to be locked before the first run and kept stable afterwards.
Write questions the way buyers actually type them, in four groups of five to eight:
- Category recommendation. "Best AI visibility tools for B2B SaaS", "which supplier should I shortlist for X". These trigger brand lists and dominate mention rate.
- Scenario. "How do manufacturers get cited by ChatGPT", "how do I market to buyers who only use AI search". Answers here lean on long-form content.
- Comparison. "A vs B", "what is the difference between X and Y". These expose how much third-party review coverage each vendor has.
- Problem-solving. "How do I check whether AI recommends my company", "where do I start with GEO". These decide who owns the methodology narrative.
Three rules. Never use internal jargon — buyers do not say "run a GEO diagnostic", they say "how do I know if ChatGPT recommends us". Log platform, language, region and date for every question. Once locked, only append: new questions get their own effective date and never overwrite the baseline.
If you have search query data, pull long-tail questions with interrogatives; that beats guessing. If you do not, ask sales to paste the last twenty opening questions from real calls. That list is usually better than anything a keyword tool produces.
Step 2: Run your brand and competitors in the same batch
Same-batch execution is what makes the comparison valid. Model versions change, retrieval changes daily, and a difference measured across two different days may be nothing more than a different retrieval context.
Practically: one batch contains your brand plus three to five competitors, answering the same prompt set, on the same platforms, on the same day. Pick two direct competitors, one category leader, and one adjacent substitute. The first two define your relative position, the leader shows the default narrative of the category, and the substitute reveals how buyers think about replacing you.
On platforms, start where your buyers actually are: DeepSeek, Doubao, Tongyi Qianwen and Wenxin for China-facing business; ChatGPT and Perplexity for global. Do not spread across eight assistants on day one. Run two or three well, then expand.
In AiVisible this is a detection project with your brand and competitor names attached to a fixed prompt set, executed as one batch and reported per brand. You can also do it by hand — same day, same questions, screenshots plus notes on order and sources — but consistency is the first thing to break.
Step 3: Read four metrics, not just "were we mentioned"
A batch produces at least four signals. Drop any one and the conclusion skews.
| Metric | How to compute | What it tells you | Common misreading |
|---|---|---|---|
| Mention rate | Prompts with a mention ÷ total prompts | Presence in the category conversation | High mention rate means recommended |
| Position | Order of appearance in the answer | First choice, backup, or "others include" | One mention equals a shortlist spot |
| Cited sources | Domain distribution behind the answer | Whether the model relied on your site, a competitor blog, or media | More citations mean better content |
| Framing | How the brand is described | Category leader, budget option, or niche solution | Positive mention equals accurate positioning |
Cited sources carry the most diagnostic value. If every citation is a review site or trade publication, your own pages never entered the retrieval pool — and no amount of title tweaking fixes that. You have to solve retrieval and extractability first.
Framing drives conversion quality. Being described as "the cheaper alternative" and "the default for manufacturing" can produce identical mention rates and wildly different pipeline. This one requires a human reading the answers; no tool scores it for you.
Step 4: Diagnose the gap — three reasons competitors get cited
The worst response to a gap is "write more articles". Find the cause first, then decide what to write or fix.
Cause one: extractable structure. Models need a claim they can lift intact. Competitor posts open with a 40–60 word definition that stands on its own, followed by an FAQ and a comparison table. To verify, open the cited page and check whether paragraph two works as a standalone answer.
Cause two: consistent entity signals. Is the brand name written one way across the site? Do Organization and Product structured data exist? Is there an llms.txt telling models which pages are fair game? Verify by searching your site for three name variants and checking whether the model resolves them to a single entity.
Cause three: third-party mention density. Trade press, directories, communities, published customer stories. Retrieval tends to weight independent sources above self-description, which is why companies with mediocre websites sometimes dominate AI answers. Check the cited domains and count how many are not yours.
For the structured data piece, work from Google's structured data documentation and the schema.org Organization definition rather than secondhand summaries. And if your robots.txt still blocks AI crawlers, start with OpenAI's crawler documentation — unblocking is the lowest-cost fix available.
Step 5: Turn the gap into a 30-day action list
A report that ends with "the gap is large" is not a report. Map each cause to an action, and keep the cycle short enough to re-run the same batch at the end of it.
- Week 1 — close content gaps. Take the prompt group with the lowest mention rate and publish three to five posts where the question is the headline. Open each with an extractable claim, close with an FAQ. Skip category overviews; answer one question per page.
- Week 2 — structure. Add Organization, Product and FAQPage markup to core pages, ship llms.txt, and standardize the brand name and positioning line site-wide.
- Week 3 — third-party presence. Trade press contributions, publishable customer stories, directory listings, partner pages. Slowest to move, most likely to decide whether you appear at all.
- Week 4 — re-run the identical batch. Same prompts, platforms and competitors, new date. Compare mention rate, position and cited sources, and write a one-page review.
One caveat worth stating plainly: timelines differ enormously by industry, platform and prompt type. Thirty days is a review cadence, not an outcome guarantee. The only thing you can guarantee is that an identical re-run tells you whether the work mattered.
Three traps that ruin the conclusion
Trap one: treating one platform as the market. Present in DeepSeek and absent in Doubao is normal, not a crisis. Read per platform, then allocate.
Trap two: drawing conclusions from a tiny sample. Going from one mention in twenty prompts to two is noise. You need 20–30 prompts across two or three platforms before a trend is worth discussing.
Trap three: confusing mention with recommendation. "Others worth considering" and "the default choice" both count as one mention and have nothing in common commercially. Position and framing have to be read together.
There is a quieter trap too: designing the prompt set in your own vocabulary. Ask "which GEO agencies are credible" and you get a vendor list. Ask "how do I get AI to recommend my company" and you get methodology and tools. The language of the prompt set decides whether you collect a competitor roster or a buyer decision path.
Make it a recurring ritual
The value of this analysis is not in run one, it is in run five. The first run gives you a gap table; by the third you can see which actions actually changed the answer. Put it on a biweekly cadence — same prompts, same platforms, same competitors, new date — and log the four metrics in one sheet.
If you have no baseline yet, start with a free AiVisible scan of your own brand to get mention rate and platform spread, then add the competitor dimension using the steps above. The baseline matters more than the method: without it, no future "improvement" can be verified.
FAQ
How often should you run a competitor AI visibility analysis?
Every two to four weeks with an identical prompt set, plus an extra run after major site or content changes. Running more often than that mostly measures retrieval noise, because the answers themselves fluctuate between runs.
Is one AI platform enough?
No, but you can phase it. Start with the two or three assistants your buyers actually use, get the workflow and metric definitions stable, then expand. A conclusion drawn from a single platform often reverses on the next one.
How many competitors should be included?
Three to five: two direct competitors, one category leader, and one adjacent substitute. The direct competitors set your relative position, the leader shows the category's default narrative, and the substitute reveals what buyers compare you against.
Can you do this without paid monitoring tools?
Yes. Fix a set of 20–30 prompts, ask them on the same day across a few platforms, and record order and cited sources with screenshots. It is slower and easier to get wrong, which is why manual runs work best for establishing a first baseline.
Can the results drive content planning directly?
Yes, once you separate two kinds of gap: no content exists on the topic, or content exists but is neither retrieved nor extractable. The first calls for new pages, the second for structural and formatting fixes.