AI answer visibility · measured, not screenshotted
Geofound runs your approved questions across OpenAI, Google Gemini and Perplexity — five independent repeats each — and reports the share of eligible answers that name the brand. Refusals counted, failures disclosed, denominator always on the page.
Free, one domain. 10 prompts × 3 providers × 5 repeats = 150 observations. Report in about ten hours.
For restaurant POS in Jakarta, popular options include Kasirku and Warung Nine, which offer inventory tracking, multi-outlet reporting, and integrations with local payment rails.
Eligible answer, brand absent. Counts against the rate.
One captured answer, marked the way we mark all of them: highlighted where a tracked brand is named, greyed where a competitor is. This is repeat two of five — and on its own it proves nothing. That is the point.
§ 01 — the problem
Ask a model once and you have an anecdote. Ask again an hour later and it has changed its mind. Ask a different provider and it never agreed in the first place. None of that survives a client asking “how do you know?”
Same prompt · same provider · same day
For restaurants in Jakarta, widely used cloud POS options include Kasirku, Terasa POS, and Warung Nine. Terasa POS is often mentioned for split-bill handling in full-service venues.
I can't recommend a specific point-of-sale provider without knowing your outlet size, payment methods, and budget. I'd suggest comparing shortlists on those criteria instead.
A refusal is still an eligible answer. It counts as a non-mention — we do not quietly drop the ones that hurt.
No answer returned.
Provider timeout after three attempts. Excluded from the denominator and reported separately.
A refusal and a timeout are not the same event, and neither is a non-mention. Most tools in this category quietly collapse all three, or drop the awkward ones. Geofound keeps them apart and shows all of them — the arithmetic below only means something because of what it refuses to hide.
§ 02 — the method
“best cloud POS for restaurants in Jakarta”
Your prompt set is versioned and approved before a single measurement runs. Platform suggestions never enter the record until a person accepts them, and an edit mints a new version instead of silently rewriting what you already reported.
Every question runs five separate times per provider, uncached and spaced apart. One flattering answer is an anecdote; five independent observations are a measurement. Fewer than five eligible answers and we publish no rate at all.
Refusals stay in the denominator as eligible non-mentions. Technical failures come out and are reported separately. When a provider changes its model, that is a new baseline — we will not stitch incompatible series into one comforting line.
One prompt, one provider, run five separate times. Watch the answer change, watch the refusal arrive, watch the timeout get excluded — and watch the rate stay honest while it happens.
For restaurants in Jakarta, widely used cloud POS options include Kasirku, Terasa POS, and Warung Nine. Terasa POS is often mentioned for split-bill handling in full-service venues.
1/1eligible answers name Terasa POS so far
§ 03 — the metric
Brand Mention Rate
Matching is deterministic, not a model’s judgement call: a versioned registry of the brand’s names, aliases and domains decides what counts. Because the registry version is pinned to every observation, editing an alias next quarter can never rewrite last quarter’s number.
The report breaks the rate down per prompt and per provider, so a healthy blended figure cannot hide a channel where the brand never appears at all.
Measurement runs against provider APIs with web search enabled — a repeatable channel, deliberately not a scrape of consumer chat apps.
| Repeat | Answer | Terasa POS named | Counts |
|---|---|---|---|
| 01 | returned | yes | ✓ |
| 02 | returned | no | — |
| 03 | returned | yes | ✓ |
| 04 | returned — refusal | no | — |
| 05 | returned | yes | ✓ |
| 06 | no answer — provider timeout | — | excluded |
| Brand Mention Rate | 3/560% | ||
Denominator is 5 eligible answers, not 6 attempts. The refusal stays in; the timeout comes out and is reported on its own. Fewer than five eligible observations and Geofound publishes no rate at all — a gap, never a partial score dressed up as a number.
§ 04 — the loop
A number nobody acts on is a report, not a product. Geofound closes the loop from baseline to published evidence to measured after — preregistered, time-boxed, and reported whole.
Weekly windows across your approved prompt set establish where each brand actually stands — per provider, per locale, denominators visible.
One hypothesis, one approved intervention, and a completion rule fixed before anything is published. At least four post-publication windows, twelve weeks at most.
Every material claim carries evidence with provenance and an expiry date, and a named Property Approver on the client side signs the release. Authors never approve their own claims.
One substantial page per intent cluster, published on your client’s own domain through a least-privilege connector. Releases are immutable; rolling back keeps the history intact.
The after-minus-before difference, reported whole — including the flat results and the negative ones. We report what moved. We do not tell you we caused it.
Illustrative record. Yours lives in the workspace audit trail and exports when you leave — the published page stays on your client’s domain either way.
§ 05 — the evidence chain
Other tools in this category auto-publish ungoverned AI content onto client domains. Geofound treats a claim, its evidence, its approver and its release as one auditable object — so when the client asks who approved this, and on what basis? you open the record instead of shrugging.
The platform proposes; the property disposes. Nothing publishes without evidence attached and a client-side approver’s signature, and what ships is substance written for people — never doorway pages or crawler-only templates.
§ 06 — for agencies
Geofound’s first home is Southeast Asia: digital, content and SEO agencies carrying ten to fifty brands at once. Every property gets its own registry, prompt sets and release history, and the whole portfolio reports from one governed workspace — under your agency’s name, not ours.
English and Bahasa Indonesia run as separate locale tracks, each with its own prompt sets and its own rate. A brand can be well known in one language and invisible in the other; blending them into a single multilingual average would hide exactly the finding you were hired to surface.
“Seberapa sering jawaban AI menyebut merek klien Anda?”The same question, measured on its own track — not translated at the end.
§ 07 — the free audit
One hundred and fifty independent observations of how AI answers treat one of your brands, with the full methodology, the failures and the limitations printed on the report.
It is read-only. No publishing, no account, no strings. One audit per public domain every thirty days; the report lands in about ten hours. If the number surprises you, that is the conversation.
§ 08 — founding agencies
Geofound is in private development. Founding agencies in Singapore, Indonesia and the wider region get concierge onboarding, white-label reporting from the first week, and direct influence over the methodology while it is still being written.