Every part of the pipeline, end to end — the one-off check, the weekly tracking on top of it, and what each metric counts. Nothing hidden, nothing hand-tuned.
When you enter a brand and category, Mentioned asks Claude to write 5 realistic questions a buyer would type when looking for a recommendation in that category — phrased the way a real person searches, not the way a marketer would ask. The questions never mention your brand name, so the test reflects an organic recommendation, not a leading one. If question generation fails for any reason, Mentioned falls back to a fixed set of neutral templates rather than guessing or skipping the check.
All 5 questions are sent to the model in parallel, each as its own independent conversation. There's no shared context between them and no hint about what brand is being tested — this mirrors what 5 different buyers asking 5 different questions would actually see.
The matcher squashes both your brand name and the answer down to letters and digits, so spacing and punctuation cannot break a match, and it folds accents — a model will write “Nestlé” and “Nestle” in the same run, and both should count. It then requires the match to land on a word boundary in the original text. Without that last rule, “abandoned” scores as a hit for a brand called “Done”.
A brand whose alphanumeric core is under two characters (“C++”) is matched literally instead. A bare single letter with no punctuation to anchor on (“X”, “R”) is declined rather than guessed at: there is no honest way to tell that from the pronoun.
When there is a match, Mentioned pulls the sentence containing it as the quoted snippet you see in your results — copied directly from the model’s output, never rewritten or summarised. The highlighting on your stored answers uses the same matcher that scored the run, so a highlight can never contradict a score.
When your brand doesn't appear, Mentioned extracts other brand-shaped names from the answer — typically bolded names or items in a numbered/bulleted list — filtered against a list of generic words so "the best option" doesn't get counted as a competitor. This is a heuristic, not a guarantee it catches every name, but it never invents a competitor that isn't actually in the text.
Your score is simply: how many of the 5 answers mentioned your brand, out of 5. That's it — there's no weighting, no confidence interval, no adjustment. 3/5 means your brand showed up in 3 of the 5 independent answers.
A one-off check writes fresh questions every time, which is right for a snapshot and wrong for a trend. If the questions change between Monday and Monday, a score moving 3/5 → 1/5 conflates “we lost visibility” with “we asked harder questions”.
So when you track a keyword — a brand and a category — its five questions are written once, at creation, and reused verbatim on every run after that. Movement in the line is movement in your visibility, full stop. It is the central design decision in the product, not an optimisation.
Tracked keywords re-run weekly on our key. A keyword can be set to daily only when your account has saved its own API key for every model that keyword asks — because a daily run spends that key rather than ours. The gate is checked when you set the schedule and again when the sweep picks the work up, so removing a key later reverts the keyword to weekly rather than quietly moving seven runs a week onto our bill.
Every one of those counts something a model actually returned, and every row links to Receipts: the stored answers it came from. Where a figure could not be measured you get an em dash and a sentence, never a flattering zero. “Nobody named you” and “nothing has run” are different findings and this product never prints one as the other.
Sentiment here is a cue count, not a model’s opinion. Mentioned takes the sentence that actually names your brand, counts listed praise and criticism words in it (handling negation), and calls the sentence positive, negative or neutral on which side has more. The whole word list is printed on the screen along with how often each cue fired and the sentence it fired on, so you can disagree with a verdict and check it in one click.
It will miss sarcasm, and it will miss praise phrased in words that are not on the list. Showing the entire classifier is the compensation. A sentiment number a model was asked to invent could not be audited this way at all.
The Actions screen runs six derivations over the runs already on your account — model blind spots, prompt blind spots, visibility drops, competitors pulling ahead, weak topics, and keywords that never ran — and ranks each finding by the gap in percentage points times the number of stored answers it spans. Every row restates a measurement that is already on another screen and links straight to it. Nothing on it is advice, and it says how many findings it cut rather than truncating in silence.
Every check spends real API budget, which is why there is no anonymous check and why a free account gets 3 checks on our key before it needs one of its own, against a 300/day global backstop. The weekly sweep — the only unattended spend in the product — is metered in real token counts against a monthly dollar budget. A keyword the budget cannot afford this month is deferred, not failed: nothing is written against it, because a red mark on a healthy keyword would be a lie about the keyword.
A request that fails in a way that can clear on its own (rate limit, timeout, provider unavailable) is retried from a pool shared across the whole run, honouring the provider’s own retry hint. A rejected key or a malformed request is not retried — it failed for a reason a retry cannot fix. And a run that does fail is recorded rather than hidden, because a visible gap beats a silent zero; the cadence clock then reads the last successful run, so one timed-out request no longer leaves a keyword idle for a week.