AI Visibility Example: 34% Citation Rate Across 4 Engines
This example shows the aiwebpageseo AI Visibility Tracker for a SaaS marketing site across ChatGPT, Claude, Perplexity and Gemini. Citation rate is 34% averaged across 60 prompts, trending +8pt over the last 4 weeks. Branded prompts at 72%, category at 41%, problem-statement prompts at only 12% — the under-served opportunity tier.
| Engine | Brand | Category | Problem | Overall |
|---|---|---|---|---|
| ChatGPT | 78% | 48% | 14% | 38% |
| Claude | 71% | 42% | 11% | 34% |
| Perplexity | 82% | 51% | 16% | 42% |
| Gemini | 68% | 32% | 8% | 28% |
| Competitor | Citation Rate | 4w Trend | Gap to you |
|---|---|---|---|
| competitor-a.com | 62% | +12pt | +21pt |
| competitor-b.com | 38% | +2pt | -3pt |
| competitor-c.com | 29% | -4pt | -12pt |
| saas.example.com (you) | 41% | +8pt | — |
Outdated pricing in ChatGPT citations
3 of 8 ChatGPT citations describe your starter plan as $19/month (correct: $24/month since Feb 2026). The fix is the live page, not the training corpus: cited answers are produced by fetching pages at query time, so a current, crawlable, unambiguous pricing page is what corrects them. Make sure the price is in the served HTML (not injected client-side), state it once rather than in several conflicting places, and mirror it in Offer schema.
Wrong-attribution mention on Perplexity
1 Perplexity response attributes your homepage tagline to a competitor. Likely caused by template similarity. Differentiate hero copy; add Organization schema with explicit name/url; submit correction via Perplexity feedback flow.
Comparative-unfavourable wording on category prompts
14% of category citations describe your product as "fewer features than [competitor-a]". Note there is no ComparisonTable type in schema.org — do not invent one. Publish a feature page that states each capability in plain, self-contained prose, mark up the genuine Q&A on it with FAQPage, and describe the product itself with Product or SoftwareApplication. If the breadth claim is simply true, the honest response is product work, not markup.
This report runs sixty prompts against four assistants on a schedule and records whether your domain appears as a cited source. That is a sample of model behaviour, not a measurement of what your buyers experience, and the difference matters enough to state plainly before anyone puts a 34% on a slide.
Assistant outputs are not stable. The same prompt can return different sources on consecutive runs, and outputs vary by account, by region, by which retrieval path the system chose and by changes the operator ships without announcing. None of the operators publishes the retrieval logic, so no tool — including this one — can tell you why a citation happened. What a tracker can honestly do is watch a fixed prompt set over time and report the direction of travel.
So: read the trend, not the number. An 8-point rise over four weeks on a consistent prompt set is a signal. A 3-point move week to week is noise, and building a narrative on it will make you chase changes you cannot attribute to anything.
Branded 72%, category 41%, problem-statement 12%. Those are not three points on one scale. They are three distinct questions with three distinct failure modes, and averaging them into a single 34% hides everything useful.
Mostly a hygiene check
When somebody names your company, your own site is the obvious source. A high rate here confirms you are reachable, coherent and not contradicting yourself. It does not indicate demand, and it cannot grow much: the ceiling is set by how often the assistant chooses to cite anything at all. The useful reading is the inverse — if branded citations are low, something is broken at the access or clarity layer, and you should go and look at your logs.
A competitive content problem
Somebody asks about the category and the assistant picks whose explanation to use. You are competing with the pages that explain the category best, which are frequently not vendor pages at all. Winning here means writing the clearest available explanation of the thing, not the most persuasive description of your product.
A content gap, not a technical fault
Somebody describes a difficulty without naming any product. The assistant answers from whatever page explains that difficulty properly. A 12% rate almost always means the page does not exist on your site — you have written about your features and not about the problem they solve. This is the largest opportunity on the report and it cannot be closed with markup, bot directives or an llms.txt file. It closes by writing the pages.
Note the pattern: the tier that is easiest to improve technically is already the highest, and the tier with the most headroom is purely editorial. That is the usual shape of these reports, and it is why "AI visibility" work that stops at configuration stalls quickly.
Three ChatGPT citations describe a price you stopped charging in February. The instinct is to think of this as something lodged inside the model that must be dislodged. Mostly it is not.
Cited answers — the ones that come with links — are generally produced by fetching pages at query time. If the answer cites your pricing page and states the wrong price, the overwhelmingly likely cause is on your side: the current price is injected client-side and the fetched HTML still shows the old one, or the old price survives on a comparison page, a landing page, a help article or a cached PDF that the retriever found first. Search your own site for the stale number before you conclude the model is at fault. It usually finds it.
Where the wrong figure genuinely came from training data, no action you take edits it. There is no mechanism you control for that, and claims that a press release will "feed the next training window" are not something anyone can substantiate. What you can do is ensure the retrievable truth is unambiguous, current, present in the served HTML, stated once, and mirrored in Offer schema so a parser cannot mistake which number is the price. When a system retrieves and the page is clear, the retrieved answer wins.
Where an assistant operator provides a feedback channel for factual errors, use it for genuine errors. It is a slow route and it is not a substitute for fixing the page.
Competitor-a is cited on 62% of category prompts and has gained 12 points; you sit at 41%. The temptation is to treat that as a gap to close with volume. The useful move is to open the pages that were cited instead of yours and read them.
What you will usually find is not a schema advantage or a backlink advantage. It is that somebody at competitor-a wrote a direct explanation of the thing the prompt asked about — with the numbers, the caveats and the trade-offs included — and published it somewhere a crawler can reach it. Assistants cite passages that can be lifted out and still make sense. Marketing copy fails that test almost by construction, because it depends on the layout, the hero image and a reader who is already interested.
The 14% of category citations describing your product as having fewer features deserves the same honesty. If it is untrue, the fix is a feature page that states each capability plainly enough that a machine cannot miss it. If it is true, no amount of content or markup will make it false, and the fix belongs to the product team. Structured data cannot manufacture a capability, and a comparison table that overstates you is a liability the first time a buyer checks.
Before any of the content advice above can pay, four things have to be true, and all four are checkable in an afternoon.
The retrieval crawlers can reach you
Not "robots.txt allows them" — that is a request, not a fact. Confirm in your access logs that OAI-SearchBot, ClaudeBot and PerplexityBot arrived and received 200s. A WAF rule matching the string "bot" produces a permissive robots.txt and zero traffic, and nothing in your application will tell you.
The substance is in the served HTML
Most retrieval systems parse rather than render. Content that appears only after a client-side fetch resolves is, to them, content that does not exist — including your price.
The markup describes the page and nothing else
Accurate JSON-LD disambiguates what your text means. It does not buy citations — there is no verified evidence that denser schema is preferred — and asserting things the page does not show is a policy violation.
Then write the answers
Self-contained passages that survive being quoted out of context. This is the only item on the list with real headroom, and it is the one that closes the 12% problem-statement tier.
Items one to three are preconditions: they remove obstacles and they are finished once they are done. Item four is the work, and it is never finished. Any strategy that inverts that ratio — heavy on configuration, light on writing — will produce a beautiful audit and a flat trend line.
Where do these citation numbers come from, and how reliable are they?
They come from running a fixed set of prompts against each assistant on a schedule and recording whether your domain appears as a cited source. That makes them a sample, not a measurement of what your buyers see. Assistant outputs vary between runs for the same prompt, they vary by account, by region and by which retrieval path the system chose that day, and no operator publishes the retrieval logic. Read a 34% citation rate as an indicator with a wide error bar, watch the trend rather than the number, and never present a single-week movement as a result.
Why is the problem-statement prompt tier so much lower than branded?
Because branded prompts have an obvious answer and problem-statement prompts do not. If somebody asks an assistant about your company by name, your own site is the natural source and a 72% citation rate mostly measures whether you are reachable and coherent. A problem-statement prompt — someone describing a difficulty without naming a product — is answered from whatever pages explain that problem well, and those are usually not marketing pages. A 12% rate there is not a technical failure. It is a content gap: you have not written the page that answers the question, so nothing of yours is available to cite.
Can I make an assistant stop repeating outdated information about my product?
You can remove the reason it is wrong, and that is usually enough over time. Cited answers are generally produced by fetching live pages at query time, so the correction lives on your site: state the current price once, in the served HTML, in a place a crawler can reach, and remove the stale statements elsewhere that are contradicting it. Where a model is repeating something it absorbed during training, there is no mechanism you control that edits it. The practical route is to make the current, retrievable truth unambiguous and easy to find, and to use the operators' feedback channels where they exist for outright factual errors.
Does share-of-voice against competitors tell me anything actionable?
It tells you which competitor's pages are answering the question you wanted to answer. That is worth knowing, and it is best treated as a reading list rather than a scoreboard. Open the pages being cited instead of yours and ask what they contain that yours does not: usually it is a direct, self-contained explanation of the thing the prompt asked about, written without marketing hedging. The 21-point swing towards competitor-a is a signal to read competitor-a's content, not a number to beat.
Should I add more structured data to raise citation rates?
There is no verified evidence that assistants preferentially cite pages carrying denser structured data, and treating schema as a citation lever will waste your time. The defensible claim is narrower: retrieval systems generally parse served HTML rather than rendering it, so JSON-LD in the initial response reaches them and disambiguates what your text means — a price is identified as a price, an author as an author. That is genuinely useful and it is worth doing accurately. It is not a volume game, and inventing schema types the page does not support is a policy violation rather than an optimisation.
Why does one Perplexity answer attribute our tagline to a competitor?
Because your hero copy is close enough to theirs that a retrieval system could not tell the two pages apart on that passage, and it picked wrong. This is a content-similarity problem with a content fix: write a homepage statement that could only describe your product, and support it with an Organization node carrying your name and URL so the entity behind the page is explicit. Misattribution of near-identical marketing copy is not a bug in the assistant so much as a consequence of a page that says nothing distinctive.
What single change most reliably improves citation rate?
Publishing pages that answer specific questions in self-contained prose. Almost every citation an assistant makes is of a passage that stands up on its own — one that can be lifted out of your page and still make sense. Marketing copy rarely survives that test because it depends on the surrounding layout, the hero image and the reader's goodwill. The pages that get cited are the ones where somebody wrote down the actual answer, in order, with the caveats included. Everything else in this report — schema, bot access, share of voice — is a precondition or a diagnostic. The content is the lever.
Related Demo Reports
Run AI Visibility Tracker on Your Own Site
Get your real AI visibility report with citation rate per engine, competitor share-of-voice and trend tracking week over week.