Suede vs Profound vs AthenaHQ: how each one samples an AI answer

A competitive intelligence analyst asked the version of this question that matters: methodologically, how do these tools differ in sampling AI answers and measuring share of voice? Not which dashboard is nicer. How is the number made.

It matters because share of voice is a fraction, and a fraction is only as good as its denominator. Two tools can both report that you appear in a third of the answers in your category and mean entirely different things, because one asked ten questions once and the other asked ten questions fifty times. The first number moves when the weather changes. The second one moves when something happens.

The question none of the marketing answers

Ask any of these tools how many times it runs a single prompt before reporting a number, and the answer is where the methodologies actually separate.

Suede's site states that it runs millions of persona-specific prompts through every major AI assistant. That is a volume claim, and volume across a whole customer base tells you nothing about the depth behind your own number. A million prompts spread across thousands of brands and hundreds of questions can still be one run per question per brand. As of this writing their public pages do not state a sampling depth, do not state a run frequency, do not publish a margin of error on any score, and do not document the methodology publicly. Their pricing is not published either.

Profound and AthenaHQ both publish prices and both make the prompt the unit you buy. Neither publishes a per-prompt sampling depth on its pricing page.

None of this makes any of them a bad tool. It makes the number they hand you unauditable, which is a different criticism and a narrower one.

Why depth is the whole argument

Language models are not deterministic. Ask the same buying question twice and you can get two different lists of brands, in a different order, citing different sources. That is not a bug in the tool and it is not a bug in the model. It is the thing being measured.

So a single run of a prompt is a coin flip you are reading as a measurement. Two runs is two coin flips. The only way to tell a real change from the noise the engines produce on their own is to run the same question enough times to see the spread, and then report the spread alongside the number.

This is why the denominator question is not pedantry. If a tool reports that you appear in four of ten answers, the useful follow-up is: four of ten what? Ten questions asked once each is a different claim from two questions asked five times each, and both are different again from ten questions asked five times each, which is fifty answers.

What each tool is built for

Profound sells prompts as the pricing unit and lets you define the basket yourself, entered manually or by CSV. If you already know exactly which questions your buyers ask, that is the most direct path from your list to a chart, and the entry tier published $99 per month when we last checked on 2026-08-10.

AthenaHQ is built around content workflows and prompt management. Writers can take a tracked query and immediately draft material against it, which is the right shape if your bottleneck is production rather than measurement. It published $265 to $295 per month on the same date.

Suede is the widest of the three in stated scope. Its site names six assistants it monitors, ChatGPT, Gemini, Copilot, Claude, Grok and Perplexity, and frames the work around corporate reputation rather than marketing alone: how a buyer, an analyst or a regulator would ask, and what the assistant says back. If your concern is what an assistant tells a journalist about your executives, that framing is closer to the problem than a keyword tool is.

Check every one of those prices against the vendor's own page before acting on it. They change, and ours is a snapshot with a date on it rather than a fact.

What Standing does, and where it is narrower

Standing asks thirty buying questions across five engines and runs each of them five times, which is 750 answers. Every score is published with the confidence interval it is accurate to, so a movement inside the band is reported as movement inside the band rather than as progress. The question basket and the sampling depth are published, so any number can be re-run by somebody else.

That design costs breadth. Five engines is fewer than the six Suede names. Thirty questions is a fixed basket rather than an unlimited one, and the free check runs a smaller basket still. A tool that tracks more prompts across more assistants will see things this does not.

The trade is deliberate. Spending the same budget on more runs of fewer questions buys you the ability to say whether a change is real. Spending it on more questions run once buys you coverage of things you cannot yet distinguish from noise. Neither is wrong. They answer different questions, and only one of them survives a sceptical CFO asking whether the number moved or just wobbled.

How to evaluate any of them yourself

Three questions, in this order, to whoever is selling.

How many times is one prompt run before a number is reported? If the answer is one, the number is a sample of size one and every week-on-week change you see includes the engine's own variance.

What is the margin of error on the score? If there is not one, the tool cannot tell you whether a change is real, whoever is looking at the dashboard.

Is the question basket published? If you cannot see the questions, you cannot check whether they are the ones your buyers ask, and you cannot reproduce the number anywhere else.

A vendor who answers all three plainly is worth more than a vendor with a longer feature list, because at the end of the quarter you will be asked whether the thing you paid for worked, and only the answers to those three questions can support a reply.