Every week, another marketing tool promises to show you where your brand "ranks" in AI search. A number goes up, a number goes down, someone in a monthly meeting says "our AI visibility improved by 12%." Sounds great. Means almost nothing.
The short version
Most AI visibility reports are too shallow because they treat one model output like a search ranking. A useful report measures whether your brand is consistently included, accurately described, cited from the right pages, and compared against the right competitors across a cluster of real buyer questions.
The metric that matters is not a single AI visibility score. It is stable, defensible representation across the prompts that shape demand.
Here is the problem nobody in the industry wants to say out loud: treating AI prompt outputs like rank positions is a category error. And the reports being sold on top of that error are giving marketers false confidence while the ground shifts underneath them.
What most people believe
The assumption baked into most AI visibility tools, and most client conversations, is that AI search works like traditional search. Submit a prompt, get a result, track whether you appeared. Up is good. Down is bad. Trend line looks healthy? You are winning.
It is the familiar comfort of a rank tracker, just relabelled for ChatGPT, Gemini, and Perplexity.
The problem is that AI search is not a rank table. It is a probabilistic system. The same prompt, run twice, can return two different answers. Run it on a different model version and you get a third. Ahrefs recently published data from over a billion AI search touchpoints showing that rankings inside AI answers do not behave like SERP positions at all. The signals that drive citation are completely different from the ones that drive rank. And Search Engine Journal's Dan Taylor documented a specific case where a tool showed client visibility "dropping" after ChatGPT changed how it formatted citation links in HTML. The optimisation had not changed. The model's output structure had.
Your report told you you were losing. You were actually fine.
What is actually happening
The volatility is structural, not incidental. Google just launched Information Agents for its AI Ultra subscribers: background monitoring bots that watch blogs, news, social posts, shopping data, and finance feeds on behalf of users, then deliver synthesised updates. Not search results. Machine-curated intelligence briefings.
When search becomes a standing monitoring layer rather than a repeated query box, "did we appear in this prompt today?" stops being the right question entirely. The right question is: are we the kind of source that consistently gets included when an AI agent is watching this market?
That is a stability and representation question. Not a rank question.
The measurement that actually matters: how often are you cited across a cluster of related prompts over time? How stable is that inclusion rate when the model changes or the phrasing shifts? What is the sentiment and context when you do appear? Are you being cited as a leader, a cautionary tale, or a footnote? Who else is being cited alongside you?
These are the signals that tell you whether you have genuine AI presence or just a lucky streak.
What this means for your business
If your current AI visibility report gives you a single score, a rank position, or a trend line that looks like an organic traffic graph, ask what is underneath it. How many prompt variants? How many model runs? How is the baseline defined? What happens to that number when the model updates?
If the answer is "we do not know" or "it just tracks citations from one tool," then what you have is not an AI visibility measurement. You have a snapshot of one tool's output on one day, rendered as a dashboard.
The harder truth is that AI visibility measurement is still early. The honest version of an AI visibility report looks more like a stability audit: prompt cluster coverage, average inclusion rate, sentiment drift, citation consistency, competitor overlap. Less satisfying than a green arrow, more accurate than one.
Foundry's view is that AI presence is something you earn through a combination of content authority, machine-readable proof, and genuine third-party citation. It compounds slowly, the same way domain authority used to. The difference is that the decay rate when something goes wrong is faster, and the feedback loops are less transparent.
Build for consistent representation. Track stability, not snapshots. And be suspicious of any report that makes AI search look neat.
The upshot
AI visibility is not a position you hold. It is a reputation you maintain. Measure it accordingly.
What should a proper AI visibility audit include?
A proper audit should include multiple prompts per intent, multiple model runs, cited page analysis, answer sentiment, competitor co-occurrence, source concentration, prompt wording variation, and a record of what changed between measurement periods.
It should also separate three different outcomes: being mentioned, being cited, and being recommended. They are not the same commercial signal.
How often should AI visibility be reviewed?
For fast-moving categories, monthly is a sensible minimum and weekly spot checks are useful around launches, press moments, major content changes, or model updates. The point is to watch drift, not panic over a single answer.
Evidence and further reading
- Bing's AI Performance dashboard for AI-generated answer citations
- Bing's expanded AI visibility insights: intents, topics, citation share, and compare
- Google's generative AI optimization guidance