Nobody buys a used car on one lap around the block.
The car starts. It sounds fine. And every reasonable buyer still asks how it starts cold, what it does in traffic, and whether there are service records. Not because the lap around the block was fake. It was real. The car genuinely started.
One start is not reliability.
A screenshot of ChatGPT recommending your brand is the lap around the block. It happened. It is also the least demanding condition your brand will ever be measured under.
The question worth answering is what happens on the four hundredth query, after the model update, when a competitor published something new last week.
Here is a way to start answering it this week, with no budget and about thirty minutes.
Why persistence is the missing measurement
The IAB’s August 2026 guidance organizes AI visibility measurement into four pillars: Presence, Prominence, Portrayal, and Persuasion. It is a real advance and it gave this market a shared vocabulary it badly needed.
Every one of those four pillars measures a moment. A response happened, and something was true about that response.
None of them asks whether it still happens next week.
That gap matters because all four can be true on Monday and false on Friday. A brand can look excellent in one scan and vanish from the same query a week later because a model updated, or because the evidence supporting the recommendation was thin enough that one new comparison article displaced it.
One recommendation is an event. Repeated recommendation is a position. You can build a strategy on the second and not the first.
Step 1: Pick three real buying questions
Not category terms. Buying questions.
“Best CRM platforms” is a category term and it will teach you almost nothing, because it is not how a serious buyer actually asks. “Best CRM for a 200-person B2B SaaS company replacing HubSpot” is a buying question, and the brands that appear in it are the brands in genuine consideration.
Write three, each reflecting a different real scenario your buyers face. Include the constraints your buyers actually mention: company size, existing stack, budget posture, industry, the problem they are trying to solve.
Step 2: Run all three across four systems
Run each question in ChatGPT, Claude, Gemini, and Perplexity. That is twelve runs.
For each, record four things. Whether you appeared. Where you appeared in the response. What category the system placed you in. And the exact language used about you.
That last one matters more than most teams expect. “A strong option for teams that need X” and “you could also consider X” are both appearances, and June 2026 research suggests they do not produce the same downstream behavior. A recommendation moved branded search 4.3 percentage points against 1.8 for a neutral mention among users with no recent engagement with the brand.
Confidence is the variable. Count appearances and you will miss it entirely.
Step 3: Vary one thing at a time
This is the step most people skip, and it is where the actual finding lives.
Take one question where you appeared and run four variations of it.
Change the company size. Add a budget constraint. Ask a follow-up question such as “which one would you choose and why.” Rephrase the problem in the buyer’s language rather than the category’s language.
Then note where you survive and where you drop out.
A brand that holds through all four variations has something durable. A brand that appears in the clean version and disappears the moment a constraint enters has a much more fragile position than a single screenshot would suggest.
Step 4: Repeat in seven days
Same questions. Same four systems. Same day of the week if you can manage it.
You are looking for two things. Whether you still appear, and whether the language about you changed.
A brand that appeared confidently last week and is hedged this week has moved, and something caused that. A competitor published. A source got indexed. A model refreshed. The change is a signal worth investigating even when the appearance itself held.
Step 5: Score it honestly
Count how many of your twelve original runs included you, and how many of your variation runs did.
There is no official scale for this, so use a plain one. If you appeared in most runs and held through most variations, you have a position. If you appeared in some and dropped out under pressure, you are near a threshold, present when retrieval happens to surface a supporting source and absent when it does not. If you rarely appeared, you have a presence problem before you have a persistence problem, and that is a different piece of work.
The brand near the threshold is the interesting case. It is not less visible on average than a settled brand with the same appearance rate. It is less settled, and those call for different responses.
What to do with the result
If you drop out under specific constraints, that tells you which buying scenarios your evidence does not yet cover. That is a content and positioning question.
If you appear on one system and not others, that usually points at source concentration rather than content quality. The system that ignores you is likely weighting source types your evidence footprint does not include.
If your language weakened between weeks, look at what changed in your category during that window. Competitor publishing, new comparison articles, and review platform activity are the usual suspects.
And if you held everywhere, note the date. That is your baseline, and a baseline is the thing that makes every future scan interpretable.
The honest limitation
Twelve runs and a handful of variations is directional data. The IAB would classify it well below decision-grade, and correctly so. It is not enough to reallocate a budget on.
It is enough to know whether the screenshot in your last deck meant anything. For most teams, that is a considerable upgrade over what they have now.
Start with the lap around the block. Then take it on the highway.
FAQ
What is an AI visibility persistence test?
It is a repeated check of whether an AI system continues to mention or recommend your brand across prompt variations, different platforms, and successive weeks. It distinguishes a one-time appearance from a durable position. A single appearance proves something happened once but says nothing about whether it keeps happening.
How do I test if ChatGPT consistently recommends my brand?
Run three realistic buying questions in ChatGPT, Claude, Gemini, and Perplexity, then run variations of each with different constraints, company sizes, and follow-up questions. Record whether you appear, where, in what category, and with what language. Repeat the full set seven days later.
How many queries do I need for a meaningful persistence test?
A manual test of roughly twelve runs plus variations is directional data, useful for establishing whether an appearance is stable but not sufficient for budget decisions. Decision-grade measurement under the IAB’s 2026 guidance requires large diverse query sets tested weekly with stated variation ranges.
Why does my brand appear in some AI answers and not others?
Usually because your supporting evidence is concentrated in source types some systems index and weight heavily while others do not. Dropping out when a constraint is added often means your evidence does not yet cover that specific buying scenario. Both are evidence problems rather than content volume problems.
What does it mean if AI language about my brand changes week to week?
It means something in your category moved. A competitor published, a new comparison article was indexed, a review platform gained activity, or the model itself refreshed. Language shifting from confident to hedged is worth investigating even when your appearance held.
Is a screenshot of an AI recommendation useful evidence?
It is useful as a starting observation and weak as proof. The IAB guidance states plainly that single-response measurement is not measurement, because AI answers vary by prompt, session, platform, and model version. A screenshot documents one favorable run without showing how many unfavorable ones surrounded it.
Start here: axissuite.ai
Axis Suite by TrendAxis