AI citation tracking: what to track, how often, at what cost

The first mistake is checking whether an assistant named your company, writing down a percentage, and moving on. A month later, the cited page may have been replaced by a competitor's page, while the percentage still looks stable.

That makes an AI citation tracking tool less useful than it looks. You need the answer itself, stored verbatim, with the cited pages beside it. Otherwise, the next report gives you a new number and no evidence for what to fix.

A score is not a record

A score is a summary. It may help you spot movement, but it is a poor audit trail.

Here's a typical case, not a specific measured one: a buyer asks an assistant to compare products in your category. The assistant names your company but cites a guide published by somebody else. That is a mention without a citation from your site. Another answer may cite your page while leaving your brand out of the prose. That is a citation without a useful mention.

You need to keep those outcomes separate. An AI citation tracker should show the following for every run:

• the question asked
• the answer returned, word for word
• every cited domain and exact page
• whether your brand, product or site appeared in the answer

The answer is the important part.

A score cannot show whether the assistant described an old plan, missed a key feature, or used a page you have since changed. Keeping the full result lets you compare the answer with the current page and decide whether the gap belongs to your content, your wider reputation, or the assistant's sampling.

The measured gap is between domains and pages

The available CiteGraph sample contains 64,667 citations across 9,791 AI answers. Those answers came from 77 scans of 44 sites in 48 categories, using 507 distinct buyer questions. The engines cited 9,192 distinct domains and 21,465 distinct pages, according to the weekly research on who owns AI answers.

That is a large sample, but it is still a sample. Your own category may look different.

In a project on our own site, 27 of the 40 domains cited three or more times in the first scan were still cited 25 days later. That is 68 percent. Of the 179 exact pages cited in that scan, 14 appeared again 25 days later, or 8 percent. Counting every cited domain, 56 of 102 came back, or 55 percent. The long tail is mostly sampling, not change.

The figure shows the practical distinction: heavily cited domains are much stickier than exact pages. A domain-level report can look stable while the page-level evidence moves underneath it.

Exact-page citations turn over quickly enough for a practical rule.

Weekly is the floor.

Store the answer verbatim

A useful tracker keeps the raw evidence before it calculates anything. Store the full answer, not a clipped excerpt. Keep the question exactly as asked. Record the engine, run date, cited URLs, cited domains and any visible answer position.

“Verbatim” means the returned wording, including the awkward wording. A paraphrase made by your reporting layer is already an interpretation. It may hide that the assistant gave two different reasons for recommending your product, or that a citation supported only one sentence in a longer answer.

The exact URL matters too. A domain can survive while the page changes. If you store only the domain, you lose the part of the result that tells you what to inspect.

Keep a snapshot of the answer with each run. Then your team can ask useful follow-up questions:

• Did the assistant change its description of the product?
• Did the same page support the answer?
• Did another page on the same domain replace it?
• Did a rival enter the answer while your citation disappeared?
• Was the result stable across repeated runs?

The tracker preserves enough context to investigate a result, rather than leaving you with a dashboard number.

Run the first useful check in under an hour

You do not need a large tracking programme to find out whether your current process is sound. Run this short procedure before you compare plans.

1. Write down the buyer questions that matter to a product or service. Include comparison, alternative and problem-led questions, rather than your brand name alone.
2. Run each question more than once on the engines you care about. Save the answer exactly as returned, including the source links.
3. Mark separate outcomes: your brand was mentioned, your domain was cited, and the exact page was cited.
4. Repeat the same set the following week. Compare the stored answers and URLs, rather than the dashboard percentage alone.
5. Open the pages that appeared and check whether they still state the facts the assistant used.

The procedure is deliberately small. It gives you a baseline for question coverage, answer variation and page churn before you buy a larger allowance.

A free starting point is the AI visibility checker. It can help you inspect the basic result, but a recurring programme needs stored answers and repeated runs. The measurement method guide explains the evidence to keep when you move from a spot check to a regular series.

Weekly is the floor

Daily tracking sounds reassuring, but it can create a pile of near-duplicate answers before you know what decision each run supports. Monthly tracking has the opposite problem: a page can disappear and return before anyone sees the useful interval.

Weekly gives you a practical minimum for most teams. It catches changes while the answer is still recent, and it gives you a repeatable point at which to review pages, rivals and questions.

The data does not prove that every category needs exactly one run each week. Your own category may be steadier or noisier, and you should measure that rather than borrow certainty from another site.

Use a higher cadence when the answer affects an active launch, a buying guide or a page you have just rewritten. Keep the question set fixed while you establish a baseline. If you change the questions, engines and site pages at the same time, you won't be able to tell which change caused the result.

Review the raw answer when a result changes. A drop in citation rate could mean your page disappeared, the assistant selected a different page on the same domain, or the whole answer took a new route.

Buy the workload, not the label

The cost of an AI citation tool depends on the unit it sells. Look for questions, runs, engines, projects and stored evidence. A low monthly price is not useful if it cannot repeat the questions you need often enough.

CiteGraph's published Starter plan is $29 a month. It includes 2 projects, 10 questions each, 3 runs per question on 4 engines, 6 tracked rivals and 30 written pages a month. Growth is $69 a month with 3 projects, 10 questions each, 5 runs per question on 4 engines, 10 tracked rivals and 60 written pages a month. Scale is $149 a month with 6 projects, 10 questions per project, 5 runs per question on 4 engines, 10 tracked rivals and 200 written pages a month. These details are on the CiteGraph pricing page.

Every plan starts with a 7-day trial, opened with a $1 card-confirmation charge, and includes 2 scans. A single scan without a subscription is $12, according to that same pricing page.

Those prices are examples of a buying shape, not a universal benchmark. Count your questions first. Then decide how many runs you need to tell a real change from one variable answer. Finally check whether the tool saves the answer and cited URL in a way your team can use later.

A free option also exists for a narrower job. The AI readiness check runs 15 checks without an account and shows the fix for each. It is useful for finding site problems, but readiness is not citation history. One tells you what may block an assistant; the other tells you what the assistant actually returned.

Make the record useful to the team

Tracking fails when the report stops with marketing. A founder needs to know whether the company appeared. An SEO needs the exact page. A content lead needs the wording that was missing. A product lead needs to know whether the answer used an old fact.

Give each finding an owner and a next check. If an answer cites a rival comparison page, save that URL and note the claim it supplied. If your domain is cited but the answer is wrong, fix the page or the source of the confusion before celebrating the citation.

Keep a change log beside the raw answers. Record when a page was published, revised or redirected. Without that context, a later improvement may look like an engine change when it came from your own edit.

The same evidence can support technical work. CiteGraph runs as an MCP server, allowing Claude, Cursor or VS Code to read a scan and apply readiness fixes in a repository, according to the MCP documentation. That is useful when the scan points to a crawl or page problem, but it does not remove the need to preserve the answer that revealed the problem.

The question to ask next is: which buyer questions need a weekly record of the complete answer and exact citation, and who will act when one changes? The answer lives in your first fixed question set, your stored verbatim scans and the page-level comparison you make after the next run.