Model surfaces

We read the surface your buyers use, not the API behind it

A model in Attensira is an assistant surface that answers your prompts, addressed by a stable channel id. Where a consumer surface can be read, that is the primary path: we visit chatgpt.com, perplexity.ai, Google's AI Overview and Gemini the way a person would. Model APIs are a labelled fallback, and the two are never averaged together.

Get Started Book A Demo

By Karl-Gustav Kallasmaa, founder of Attensira · Last updated 3 September 2026

In short

Five stable channels — ChatGPT, Perplexity, Claude, Gemini and Google AI — answer your prompts on a daily cadence. Every run records whether it came from a consumer surface or an API, and the two are never blended. A channel you have not configured returns not tracked rather than zero, and a surface with no reader returns not readable.

Consumer first

The primary path reads the surface a buyer actually types into. The API is a labelled fallback where no consumer surface can be read.

Stable channel ids

When a vendor swaps the model behind their product, the channel id does not change, so your history stays continuous.

Tracked and readable

Two flags on every per-model figure. Neither one is a zero, and a dashboard that draws them as zero invents a failure.

Never averaged

A window spanning both paths reports as mixed rather than silently blending a consumer reading with an API one.

The problem it solves

Asking a model through its API is not what someone sees when they type into chatgpt.com. Different model version, different system prompt, different retrieval. A visibility number built purely on the API answers a question nobody asked, and it will disagree with what your own team sees when they check by hand — which is the fastest way for a measurement product to lose an argument it should win.

Attensira reads the consumer surface as the primary path and labels every run with the path it came from. That is why Google's AI Overview is measurable here at all: it has no API, so consumer reading is the only way to see it.

How it works

In code and on the wire a model is called a channel, and the channel id is deliberately stable. You are tracking the surface a buyer uses, not a specific model version whose name they will never see. The readable set is defined in the engine and is not the same as the list of assistants that exist in the world:

Channel	Wire id	API reader	Consumer reading
ChatGPT	chatgpt	Yes	Yes
Perplexity	perplexity	Yes	Yes
Claude	claude	Yes	No
Gemini	gemini	Yes	Yes
Google AI	google_ai	No API reader	Yes

A new workspace starts on two channels, ChatGPT and Perplexity, and you add the rest yourself. Adding a channel does not backfill: its history begins the day you add it, exactly like a new prompt. Turning a channel off stops future runs without deleting the readings already collected, and the gap it leaves is reported as not tracked rather than as a drop.

Each prompt, model and country combination runs at most once per workspace-local calendar day, at your plan's sampling depth, with the runs issued concurrently on the deeper plans so that all of them see the same model state. Failed runs are stored but excluded from n, so a channel having a bad day shows up as a smaller sample rather than as a fall in your rate.

Every per-model figure carries two flags, and reading them wrong is the most common way to misreport a result:

What you see	What it means
tracked: false	Not configured here; never queried. Not a zero.
readable: false	No reader exists for this surface. A property of the platform, not a setting. Not a zero.
value: null, n: 0	Tracked, but no successful run in the window. Not measured.
value: 0, n: 12	Asked twelve times, named zero times. A real zero.

A worked example

A workspace on the default two channels adds Claude and Google AI in the same week, then reads a per-model breakdown seven days later. Four rows come back and only two of them are comparable.

ChatGPT and Perplexity carry a full window of runs, so their rates can be read against each other and against the previous window.
Claude and Google AI carry a week of history and no more, because adding a channel does not backfill. Their deltas will mostly return no proven change until n builds.
Google AI's coverage comes from consumer reading only — there is no API reader for it — so expect thinner, less regular data there, and say so out loud in any report built on it.
Gemini, if it is not in this workspace's channel set, reports not tracked. That is not a zero, and it is not evidence about Gemini.

Deciding how wide to go is a real trade-off, because every channel multiplies the work: each one adds a daily reading per prompt, per country, at your plan's sampling depth. Wide and shallow — four or five channels over a smaller prompt set — is right when you do not yet know where your buyers are. Narrow and deep — two channels over many prompts — is right when you already know one surface drives your pipeline and you want statistically usable rates across a wide question set.

Before adding a channel across the board, add it and wait two weeks. Rates on a new surface are noisy until n builds, and the two-proportion z-test at 95% that guards every delta will correctly refuse to call a change until it has the evidence.

What it does not do

It does not track every assistant in existence. Five channels are readable, and a surface Attensira cannot read is reported as not readable rather than as an empty chart.
It does not backfill. A channel added today begins its history today.
It does not average paths. Consumer and API readings are labelled, never blended; a window spanning both reports as mixed.
It does not observe real users. Attensira asks your prompts on a schedule. It does not see the questions real people asked, or how many of them there were.
A model is not a crawler. A model answers a prompt; a crawler fetches pages. A high crawl rate does not produce mentions, and a model can name you having never crawled you — see crawler logs.
An agent model is a different concept entirely: that is the LLM a chat inside Attensira runs on, and it has nothing to do with the surfaces measured here.

One more honest edge, because it is the kind of thing a plan comparison usually hides: the channel set is not identical on every plan. Four channels are available on all of them; Gemini is the one surface Starter does not read, and it unlocks on Growth. The table below is generated from the same ladder the billing system uses, so it cannot drift away from what you are actually sold.

Which plans include it

Every plan reads ChatGPT, Perplexity, Claude and Google AI. Gemini is included on Growth and Business. What your plan otherwise governs is prompt slots, sampling depth and retention — the volume of evidence, not permission to ask.

Plan	Per month	Prompt slots	Sampling depth	Model surfaces	Retention
Starter	$129	50	n=1	ChatGPT, Perplexity, Claude, Google AI	3 months
Growth	$349	150	n=3	ChatGPT, Perplexity, Claude, Gemini, Google AI	12 months
Business	$699	350	n=3	ChatGPT, Perplexity, Claude, Gemini, Google AI	Unlimited

Yearly billing is charged at ten months' price. Full plan detail is on pricing.

Frequently asked questions

Do you read the consumer app or the API?

The consumer surface first, wherever one can be read — chatgpt.com, perplexity.ai, Google's AI Overview and Gemini, the way a person would. Model APIs are a labelled fallback used only where no consumer surface can be read. Every run records whether it was consumer or API, and the two are never averaged; a window spanning both reports as mixed.

Which models does a new workspace track?

Two: ChatGPT and Perplexity. That is a starting configuration rather than a limit — you add the others yourself in settings. A channel you have not configured is never queried, and its metrics come back as not tracked rather than as a zero.

Are models gated by my plan?

Almost not. ChatGPT, Perplexity, Claude and Google AI are available on every plan. Gemini is the exception: it is not part of the Starter channel set and unlocks on Growth and Business. The plan table on this page is generated from the same ladder the billing system uses.

Why is a model showing nothing instead of 0%?

Because nothing is a different fact from zero. Not tracked means the channel is not configured for your workspace, so it was never asked. Not readable means no reader exists for that surface at all — Google AI over the API is the live example. A measured zero is a value of 0 with an n above zero: the model was asked and did not name you.

Will adding a model fill in past data?

No. A newly added channel starts producing data from its next scheduled run and does not backfill, so its history begins the day you add it. Turning a channel off later leaves a not-tracked gap in the series rather than a drop, and does not delete readings already collected.

Is a model the same thing as an AI crawler?

No, and Attensira reports them as separate systems. A model answers a prompt; a crawler fetches pages from your site. They also fail differently: AI crawlers do not execute JavaScript, so crawler traffic has to be measured in edge middleware rather than with a client-side beacon.

You set the destination. It handles the route.

Attensira measures what assistants answer, finds what is holding you back, and opens a pull request you review before anything ships.

Get Started Book A Demo

We read the surface your buyers use, not the API behind it

In short

Consumer first

Stable channel ids

Tracked and readable

Never averaged

The problem it solves

How it works

A worked example

What it does not do

Which plans include it

Frequently asked questions

Do you read the consumer app or the API?

Which models does a new workspace track?

Are models gated by my plan?

Why is a model showing nothing instead of 0%?

Will adding a model fill in past data?

Is a model the same thing as an AI crawler?

Keep reading

Prompt tracking

Mention rate

Citation tracking

Crawler logs

You set the destination. It handles the route.

The problem it solves

How it works

A worked example

What it does not do

Which plans include it

Frequently asked questions

Do you read the consumer app or the API?

Which models does a new workspace track?

Are models gated by my plan?

Why is a model showing nothing instead of 0%?

Will adding a model fill in past data?

Is a model the same thing as an AI crawler?

Keep reading

Prompt tracking

Mention rate

Citation tracking

Crawler logs

You set the destination. It handles the route.