Branded recall is not discovery: what our DoneHat self-audit recorded

Historical evidence and scope

On 2 October 2026, DoneHat appeared in all nine usable branded checks in our saved self-audit. The same audit recorded no mentions in eleven usable unaided checks. Asking about our name and asking for a solution produced different recorded results.

The result is directional: the positive calibration control failed, so mention detection may have undercounted. Treat the counts as a starting point for reviewing the answers, not proof that an assistant never recommends us.

Three question types, three denominators

Unaided discovery — 0 / 11

Comparison inclusion — 0 / 5

Branded recall — 9 / 9

An unaided question asked, “Which AI visibility tools show measurable results for teams?” A branded question asked, “What does DoneHat do and who is it best for?” The first asks the assistant to select candidates. The second supplies a candidate. Combining their results would hide that distinction.

These denominators count usable answer checks. They are not people, Google impressions or market share. Repeated questions are clustered, so the checks should not be treated as an independent survey of buyer demand.

Make the next check useful

Start with the decision: do buyers encounter us while asking for a solution, and do answers about our name describe us accurately?

Keep unaided and branded questions in separate groups. Save the exact question, observation time, engine or model where available, answer text and cited sources. Record usable, failed and not-run checks separately before calculating a rate.

Then review the answer behind the result. A missing mention can be worth investigating, but a citation does not establish why the assistant chose or omitted a brand. A failed detection control needs review before the headline becomes a strong claim about absence.

The saved core receipts reconcile to 27 attempts: 25 usable checks and two failures excluded from the rates. Their usable engine labels were ChatGPT and Perplexity Sonar. The report also retained six separate AI-generated answer samples. Its broader methodology count is 31: 25 usable core checks plus those six additional samples. That is a different population from the 27 core attempts, and it is not the denominator of the three rates above.

Keep each audit dated

A read-only review on 23 October also found a later own audit, measured 4 October at 18:20 UTC. It recorded 0/9 unaided, 0/5 comparison and 9/9 branded mentions. Its 24 core attempts produced 23 usable checks and one failure, and it was marked directional with a calibration warning.

These are two historical samples. We have not established a matched question set or controlled publication between them, so we do not treat their counts as a before-and-after improvement test. The later audit is not a measurement of current October 23 visibility.

Use search questions as context

A later saved Search Console export covered 23 September–20 October 2026 and was retrieved 22 October. Its returned query row for “how to measure brand visibility in chatgpt” recorded 20 impressions and zero clicks. That is an editorial clue: write a practical explanation for someone trying to measure visibility.

It is not a count of questions asked inside ChatGPT. Nor does it establish a link to the 2 October audit: the audit had no Search Console data connected and inferred its questions from the site and category. The two sources have different dates, units and coverage.

Change a fact you can check, then retain the evidence

Once you have reviewed the answers, choose a change the evidence supports. You might clarify a product fact, repair an accessible page or answer a buyer's unresolved question. Keep the reviewed change and checked live URL together.

A later recheck should preserve the question and enough measurement context to explain changes in the test itself. Report unchanged, regressed or inconclusive results as clearly as positive ones. This historical example contains no matched after measurement and establishes no ranking, revenue or publication-caused improvement.

Evidence and method

Measured 2 October 2026 at 18:11 UTC, own DoneHat domain only. Selection: all saved individual trial receipts, grouped by their recorded question mode; only usable receipts enter mention-rate denominators. Counts: unaided 0/11, comparison 0/5, branded 9/9; 27 attempts, 25 usable, 2 failures excluded.

Quality on 2 October: directional. The positive control failed and mention detection may undercount. The 4 October audit also carried a directional label and calibration warning. Questions repeat; no population-wide estimate, independent market-share claim or matched before-and-after conclusion is made.

Population definitions: 2 October core attempts were 27, consisting of 25 usable and two failed checks. The broader methodology count of 31 includes six additional single-answer samples alongside the 25 usable core checks; it does not replace the core rate denominators. On 4 October, core attempts were 24, with 23 usable and one failed check.

Original saved audit receipt SHA-256: 56e22a12760d7773355d59aac54df9d9a4d8914b524b374b8c29baa67b988689. This checksum identifies the source file reviewed; it is not independent validation of the classifications.

The Search Console context is one historical returned query row, from a domain-property export with independent query/page responses; no query/page join is inferred. Original saved Search Console receipt SHA-256: d9fc8cd9ff960cc4e519953d31b83a703e68e12bd958603f5d4e1ccbd992ad15.

Read-only retrieval reviewed 23 October 2026 at 21:37 UTC. The saved own-domain extract includes the historical 2 and 4 October reports; extracted report SHA-256: 7c58987535c4fc368192da41162b78be07bc4241649a3157187d551a5134e528. New retrieval does not turn an old measurement into current performance.