Skip to content
Coverage

Six surfaces, measured differently, declared honestly.

Every engine below is included in every plan. They are not, however, the same kind of measurement — two are reached through a search vendor, four through grounded model APIs, and each has a failure mode that changes how its figure should be read. Here is all of it, including which adapters are sourcing answers today and which are still in build.

00Adapter status

What is measuring today.

Two different facts get confused with each other on pages like this, so they are separated here. Bundled is commercial: every engine is in every tier, permanently, with no per-engine add-on. Live is operational: which adapters are sourcing answers right now. The first does not move. The second does, and it moves in one direction.

Adapter status · as of 2026-08-271 of 6 sourcing answers
ChatGPTIn buildGrounded-API adapter, built against the same interface as Claude.
ClaudeLiveSourcing answers in production, verified end to end through the provider's batch interface.
GeminiIn buildGrounded-API adapter, routed through Vertex — the consumer developer API exposes no location control, and an unlocated South African figure is not one we would print.
PerplexityIn buildGrounded-API adapter.
Google AI OverviewsIn buildBlocked on the South African vendor bake-off: three SERP vendors benchmarked on ~100 ZA queries for AI Overview detection rate, AI Mode fidelity and cost, before either Google adapter is written.
Google AI ModeIn buildSame bake-off. Detection quality on these surfaces varies enough between vendors that choosing one on price would corrupt the trendline for a year.

This table is generated from the same constant the monitoring loop reads to decide which engines it can query. There is no second list to keep in step, so the page cannot quietly stay flattering after the product changes — which is the only version of a status page worth publishing.

A cycle running against fewer engines is not a broken cycle. It reports the legs it reached, the bands widen to match the smaller sample, and the coverage line on the report names what is missing — the same machinery that handles a vendor outage handles an adapter that has not shipped yet. See how six legs become one number.

01Why bundled

The per-engine add-on is a margin trap.

It is the category’s standard pricing structure: two or three models in the headline price, the rest sold separately — roughly €20–30 per engine at one competitor, and $7 to $115 a month per engine depending on tier at another.

For a single brand it is a line item. For an agency running twenty clients across six engines it is a compounding cost that arrives after the client contract is signed, and it makes “which engines do you cover?” a question with a commercial answer rather than a technical one.

So all six are bundled, and the unit economics were solved first — batch submission and a mid-tier analysis model cut cost per run by 79% with the mention rate unchanged. A bundle whose costs were never modelled is a bundle that gets quietly degraded in month eight.

How the economics work

Not included, on purpose

  • Microsoft Copilot. Minority coverage across the field, add-on-only where it is offered at all, and no demand signal from this market. Available if a market asks.
  • Consumer-interface scraping. Higher fidelity, terms-of-service exposure your client would inherit, and a cost base that does not survive channel pricing. A calibration tier is on the roadmap; a scraping-first product is not.
  • Engines we cannot geo-target. Not excluded — but where a market cannot be located, the figure says so instead of pretending.
01Assistant

ChatGPT

GPTIn build

How we reach it

Grounded model API with web search enabled

Location control: User-location parameter on the request

Its specific failure mode

Search does not always fire. In published testing on identical prompts, roughly 58% of API runs never performed a web search at all — so an absence can mean the engine never looked rather than looked and found nothing.

How to read this engine’s figure

Every capture records whether search fired. Runs that did not search are visible and can be excluded, rather than silently depressing a mention rate.

02Assistant

Claude

CLDLive

How we reach it

Grounded model API with the web-search tool

Location control: Country, city and timezone on the search tool

Its specific failure mode

Location control is the most complete of the assistants, and there has been an open report of incorrectly localised results — so it is validated rather than assumed.

How to read this engine’s figure

Our South African smoke tests compare located against unlocated runs on local-intent prompts, and the comparison is repeated when the adapter changes.

03Assistant

Gemini

GEMIn build

How we reach it

Grounded model API via Vertex

Location control: Coordinates through Vertex only — the consumer developer API exposes none

Its specific failure mode

The straightforward developer API has no location control whatsoever. A tool using it is producing an unlocated figure and, unless it says so, presenting it beside five located ones.

How to read this engine’s figure

We route through Vertex for geo-grounding and label the method on the capture. Where a market cannot be geo-grounded, the figure carries the limitation rather than hiding it.

04Assistant

Perplexity

PLXIn build

How we reach it

Grounded model API

Location control: Full location filters

Its specific failure mode

The most citation-dense surface we measure, which makes it the most useful engine for source analysis and the least representative for answer length.

How to read this engine’s figure

Reported on its own leg as well as in the aggregate, because a citation-heavy engine can dominate an unweighted source table.

05Google surface

Google AI Overviews

AIOIn build

How we reach it

Third-party SERP vendor

Location control: Vendor geo targeting, plus device

Its specific failure mode

Detection is imperfect. Published benchmarks put best-in-class AI Overview detection near 68% — about a third of genuinely present overviews return nothing. Trigger rate also varies enormously with intent: around 36% on informational queries against 5–8% on transactional ones.

How to read this engine’s figure

Presence is tri-state — present, absent, not detected. Undetected observations are excluded from rates and counted on the report, so a bad vendor week widens a band instead of inventing a decline.

06Google surface

Google AI Mode

AIMIn build

How we reach it

Third-party SERP vendor

Location control: Vendor geo targeting, plus device

Its specific failure mode

Live in South Africa since August 2025 and, since March 2026, answering in Afrikaans, isiZulu, Sesotho and Setswana. Costs roughly twice an organic SERP call to retrieve and takes appreciably longer.

How to read this engine’s figure

Included in every tier despite the cost, because it is the surface most likely to become the default way South Africans search.

07Aggregation

How six legs become one number.

The aggregate is a market-weighted roll-up, not a flat average. An engine nobody in the market uses should not carry the same weight as one everybody does.

StepWhat happens
Per-engine legEach engine's own band and run count, computed only from its own runs
Coverage checkLegs that did not complete are excluded and named, rather than being treated as zeroes
Market weightingLegs are weighted by the market's engine mix — local usage, not global averages
Aggregate bandA binomial interval over all contributing runs, reported with total n and coverage

A six-engine aggregate and a four-engine aggregate are different measurements, and the product never presents one as the other. When coverage drops, the band widens and the report says which legs are missing.

Next step

See it against your own client list

A working demo runs your prompts, in your market, on live engines — not a sandbox with seeded data. Bring one client brand and three competitors.