questions worth answering straight

Weighted by three things, on every brand mention found in an answer: how prominent the mention is (first in a list scores far higher than tenth, using logarithmic decay, which mirrors how people actually read a list), how positive or negative the mention is (a recommendation counts fully, a hedge counts partly, a dismissal counts as zero, but still gets recorded as a finding), and how commercial the question was (a shortlist question is weighted at full value, a self-description question is excluded from share of voice entirely because it measures accuracy, not share). The weighted number, a raw appearance count, and a coverage ratio are always reported together, never the weighted number alone, because the weighted figure is the insight and the raw figure is the receipt anyone can verify by counting.

From a measured run on the brand Linear: it appeared in 3 of 4 engines, with a weighted share of voice of 0.12, well behind the category leader Jira at 0.236. The same run surfaced ClickUp as a competitor nobody had entered manually. That is the kind of finding a percentage alone would not have surfaced.

We match whole words and domains, never substrings, specifically because of this. Measured directly: a provider index returned 2,420 mentions for the keyword "Unbound" while the domain withunbound.com returned zero on the same query, a three order-of-magnitude gap driven entirely by the word being an ordinary English word as well as a brand name. Brands flagged as common-word risks get strict matching, where the name only counts when a domain or a category term appears near it, and any match below a confidence threshold is shown with its snippet and an uncertainty marker rather than silently promoted to full confidence.

Because it isn't one. A call the provider could not complete is excluded from both the numerator and the denominator of the ratio, and it lowers a separately reported coverage figure instead. Folding a failure into the zero column would quietly punish a brand for an outage that had nothing to do with its visibility. A related case: an answer that names nobody at all is also not treated as a zero. It means that particular question did not discriminate between vendors, not that the brand lost, so it is excluded from the calculation and counted separately as non-informative.

A single measurement dressed up with a precision band is exactly how credibility gets spent. A band needs at least three runs of the same prompt set, market and weights within the measurement window before it is shown at all; below that, the point value is shown with an explicit note that there is not yet enough data to estimate variability. Once there is a band, two snapshots are only reported as a real, reportable change when their intervals do not overlap. If they overlap, the product says the movement is not distinguishable from normal noise, on purpose, because telling a customer they moved when that cannot be shown is how the whole method loses the trust it depends on.

No. Detecting whether your brand appears is deterministic text inspection over the stored final answer, whole-word and domain matching with a declared confidence tier for each method, not a model being asked to judge the answer. A model guessing at whether you appeared would be fabrication dressed as measurement, which is the one thing this product is built not to do.

Yes, and the exclusion is counted, not silent. Reasoning models return several items in one response, and the ones marked as internal reasoning are the model's private draft, not the answer a buyer sees. That draft names brands the final answer then drops. Counting it would report a brand as visible somewhere no buyer ever looked. Every stored run records how many reasoning items were excluded, so the exclusion itself is auditable rather than invisible.

You get a partial result with its coverage stated plainly, never a number quietly padded to look complete. Below 70% coverage, any aggregate share of voice is shown with a prominent warning. Below 50%, no aggregate is shown at all, only the individual cells that did complete, with an explicit statement that there is no honest basis for combining them into one number. Nothing about the missing cells is invented to fill the gap.

Only when they are actually comparable, and the product checks this rather than assuming it. A prompt set version change, a different market or language, a different mix of engines, a different weight per engine, or a different model within the same engine all break the comparison, because any one of them changes what the denominator means. When that happens the comparator names the difference instead of drawing the line anyway, and can offer to renormalise over the engines the two runs share in common, while saying plainly that it did.

Because the question is the same question. "Best growth marketing agencies for AI startups" asked on behalf of one company is the identical prompt, engine and market as the same question asked for another company in the same category, so the expensive live call is run once and the answer is stored once. What never gets shared is the reading of that answer: your brand, your competitor list and your results are computed privately over the stored text and carry your account identifier, enforced at the database level. The date the answer was captured is always shown, which is what keeps sharing honest rather than a way to quietly serve stale data as fresh.

Engines, overwhelmingly. A live call to the cheapest engine we measure and a live call to the most expensive one differ by roughly 3.8x, and that gap comes from the model and the token budget behind each engine, not from anything about the request itself. Prompt count moves the bill in a straight line and by comparison very little. That is why the tiers are separated by how many engines they check and whether prompts are shared or your own, not by prompt count alone.

They exist because on-demand usage, not the scheduled prompts, is what erodes margin. On a mid-tier plan, nominal scheduled usage left a wide margin; adding ten extra on-demand runs in a month dropped it noticeably, and running the same plan daily instead of weekly dropped it further still. An MCP-connected agent can call on-demand runs in a loop with nobody watching a screen, so the quota is a hard limit enforced on every call, including read-only ones, not a soft target applied after the fact.

Yes, from the account settings screen, at any time, without asking us for it. The export is a single file containing every brand, competitor, market, prompt set, scan, cell, stored answer, cited source, mention, metric, finding and diagnosis tied to your account. The data belongs to you, so leaving with all of it should be easy, not a support ticket.

Cancellation is handled through Stripe's own billing portal. Access does not stop mid-cycle: it runs through the period you have already paid for, and your account reverts to the free tier once that period ends. If a payment fails, Stripe retries automatically before anything about your access changes.

No. Four processors handle specific, limited pieces of the system: a database and sign-in provider, a hosting provider, the API provider that relays questions to the AI engines, and a payments provider if you subscribe. Each gets only what it needs to do its job. There are no advertising trackers and no third-party analytics anywhere on this product.

It is optional, used only so we can send you the link to your own report, and it is not required to run the scan or see the result. If you gave one and want it removed, write to us with the domain you scanned and we will delete it.

Not in a readable form. A one-way hash of it is kept only to rate limit the free scan, so the same visitor cannot run an unlimited number of scans, and that hash is not reversible back to an address.

Asking it yourself is a reasonable first gut check, and it is free. What it will not give you is a repeatable measurement: your own session carries your account history and your location, so you are not seeing what an anonymous buyer sees, and there is no way to tell a real change from ordinary variation between two of your own attempts. A scan here runs the identical prompt set, geolocated to a market you choose, with no account history, and produces a dated report at a permanent link you can forward, rather than a screenshot.

A classical SEO platform measures rankings, keywords and backlinks. This measures whether an AI answer engine names your brand at all, and what it read to decide. Those are different questions with different answers, which is why a strong SEO position and poor AI-answer visibility can coexist. A basic classical SEO layer is included here too, since it costs a small fraction of what an AI-answer measurement costs and there was no reason to leave it out.

The material differences, where we have checked them, are that the full answer text is the primary output here rather than a metric on its own, and that the scoring method, its weights and its confidence rule are published in full rather than kept as a proprietary black box. See the comparison pages for the specific, named differences against each one, including where they are honestly the better choice.