Should Your SEO Company Track AI Visibility?

Normal SEO reporting leaves part of the answer unseen

Ordinary SEO reporting doesn't answer this question in full. Traditional SEO reports cover rankings, impressions, and clicks, but searchers increasingly get their answers directly in AI assistants. So, metrics based on visits to your website or clicks from Google's search results miss much of your reach.

Google now provides a Generative AI performance report in Search Console. It can show impressions and the pages involved in Google's generative results, broken down by country, device and date. That is useful evidence from Google, but it does not show whether ChatGPT mentioned your company or which competitor another assistant recommended.

OpenAI referral links automatically include `utm_source=chatgpt.com`, so clicks from ChatGPT to your website are measurable in your analytics as referring traffic. But this does not count people who read an answer without clicking. A saved answer can show whether the company was mentioned, recommended, cited or described incorrectly, but it is an observation from that product under those conditions, not a permanent rank.

Your SEO company therefore needs more than one data source. Search Console can show Google AI exposure. Site analytics can show visits that reached the website. Saved answer checks can show what a potential customer saw before a click. Calls, forms, booked consultations and accepted jobs remain business outcomes. Those four layers should not be blended into one number.

One answer is not a trend

Generated answers can change between otherwise similar runs. AnswerPrism reanalyzed the public MIA-FR dataset, which recorded 295 valid answers from 25 non-branded questions tested across four AI products with three planned repetitions. Among the 100 question-and-product groups with at least two valid runs, the business's mention status changed within 15 groups. Citation status changed within 11.

But there are reasons a single screenshot and one blended score, as shown by an AI visibility checker, do not make much evidence. This is based on our reanalysis of one dataset of 295 observations, which concerns one French agency under one dated test design. Your results will likely differ, but this dataset does serve as evidence that repeated runs over time and separate fields for mention and for citation are important. A single screenshot misses the run-to-run variance; blending unlike signals from mention and citation into one score misses each separately; and showing only the latest answer misses the prompt list and number of runs.

This is not a local-business benchmark. It is a bounded reason to keep repeated runs and separate outcomes visible. That does not mean an agency needs to run hundreds of questions every day. A smaller fixed set of real buyer questions, repeated consistently and preserved in full, is more useful than a large changing set chosen to make the trend look active. The broader measurement guide explains how to keep the questions and conditions comparable.

Ask to see the evidence each month

The owner should not have to accept a score on faith. The monthly report should name the products checked, the exact questions, the location or market, the dates, the search mode when visible and the number of runs. It should preserve every valid result, including zero mentions, rather than show only the best answer.

At minimum, ask the SEO company to keep these results separate:

A score can summarize that evidence, but it should not replace it. If the agency uses a proprietary tool, ask what the denominator is, which questions were included and whether you can inspect the raw answers. A trustworthy AI visibility score should be reconstructable from its prompt set and counting rule.

The report should then state what changed, what evidence supports the observation, what action is proposed, who owns it and when the same conditions will be checked again. A change in an answer after a page update is worth noting. Unless the work isolated that change from everything else, it is not proof that the edit caused the result.

• Repeated runs over time, not just a single run
• Separate fields for mention (brand name or your website address) and for cited pages (your website address or other URLs)
• A clear listing of prompts tried, not just a single one
• Recommendations, named competitors and factual errors in the answer
• Visits from known AI referrers, Google AI impressions and qualified enquiries

Keep the work, traffic and business result separate

An agency can repair an inaccessible page, clarify a service, correct an address, improve internal links or add proof to a thin page. Those are inspectable pieces of work. A later mention may be encouraging, but the agency does not control the assistant's answer and should not promise a placement.

A useful report follows the path without collapsing it. First, record the answer-layer observation: who appeared, which page was cited and whether the facts were right. Second, report measurable exposure or visits from Google and known AI referrers. Third, connect those visits to qualified calls, forms, consultations or sales where the evidence allows it. A citation is not a visit, and a visit is not a lead.

There is no claim that any single one of these changes is sufficient to ensure inclusion. The report can say that an answer changed after the work. It should not quietly turn correlation into causation.

Most of the work belongs inside modern SEO

If the same provider already manages the website, technical search work, content, internal links and major local profiles, much of the underlying work is already SEO. A service page does not become a different deliverable because an AI assistant may read it. The same is true of correcting public business facts, making useful proof visible and removing a crawl barrier.

The owner should not pay twice for the same underlying task simply because it has been relabelled. Ask the agency to separate new measurement work from website or profile work already covered by the agreement. A wider test may justify additional time; renaming an ordinary page edit does not.

A separate project, or perhaps a separate specialist within the existing provider, may be appropriate to address a wider multi-platform baseline, a measurement protocol to be followed consistently, or substantial off-site research; or for work that the provider cannot perform. It can also make sense when nobody currently owns the measurement or when the existing agency cannot preserve raw answer evidence.

Before approving that extra scope, ask what the project will add: another platform, a repeatable baseline, source research, competitor analysis or work outside the current provider's capability. The answer should identify a missing job, not merely a new label.

Make the report earn its place

At the next review, ask what decision the AI visibility section changed. Did it reveal a wrong fact that needs correcting? A customer question the site does not answer? A competitor source worth investigating? A useful page that is being cited but not converting? If the agency cannot connect the evidence to a defensible action, the report is not finished.

It shouldn't be hidden. Normal variation, failed runs and unchanged results belong beside the improvements. But they don't promise a ranking, a lead, or revenue.

A local business does not need another dashboard to admire. It needs enough evidence to decide whether to correct a fact, improve a page, investigate a source, change the test or leave the work alone. Pay for the report when it helps make one of those decisions.