Measure who AI answers cite, as an API instead of another dashboard

The answer layer moved, and reporting has not caught up

A query now resolves in several places at once: ten organic results, an AI Overview with its own citation set, AI Mode, and whatever the assistant your customer actually uses decided to say. Those surfaces disagree, and the disagreement is the finding.

Most teams try to reconstruct this by hand — one tab per engine, a screenshot, a spreadsheet on Friday. It does not survive contact with a second market or a second week, and it cannot be diffed.

• A brand can hold position one and be absent from every answer above it.
• A citation is a domain, so share of answer is countable — as long as something counts it the same way every time.
• If a percentage comes out of a language model, you cannot reproduce it, and a number you cannot reproduce does not belong in a board deck.

One call, every surface, one timestamp

1. Step 1

   The surfaces run together

   Organic results, the AI Overview and its citations, optionally AI Mode, and a set of answer engines are collected for the same query at the same moment — so what you compare is one query, not one query per afternoon.

2. Step 2

   The counting is deterministic

   Share of answer per domain, agreement between engines and the points where they diverge are computed in code from the collected citations. The same response always produces the same numbers; a model writes the reading, never the arithmetic.

3. Step 3

   A missing surface is reported, not faked

   If one engine is unavailable the call still succeeds and says so with partial: true. Silently returning a smaller denominator would move every percentage on the page.

The call, in full

curl -X POST https://api.dataswap.io/v1/geo/answer-landscape \
  -H "Authorization: Bearer $DATASWAP_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "best home ev charger",
    "engines": ["engine-a", "engine-b"],
    "include_ai_mode": true,
    "gl": "us"
  }'

The calls, and what they cost

A composed call quotes a ceiling built from the operations it may run — the search, the AI Overview, AI Mode, one call per answer engine — and settles on the ones that actually ran. Fewer engines is a cheaper call, and the receipt shows which parts were charged.

1 credit = $0.002 on every plan. Full plans and prepaid packs are on the pricing section.

What this is not

• It is not a scheduler. Each call is one measurement of one query; the cadence, the storage and the alerting are yours to build — which is the point of buying the layer rather than the dashboard.

• It is not a rank tracker. Scheduled keyword and local rank tracking are not part of the product today, and we would rather say so than sell a cron job you have to write anyway.

• It samples answer engines, and a sample is not a census. An assistant can answer the same prompt differently for two people; treat a single reading as one observation.

• It does not tell you why you were cited. Correlating citations with your own pages is analysis we do not do for you.

Questions we get asked

Are the percentages produced by a language model?

No. Share of answer, agreement and divergence are computed in code from the citations that were collected, which is why the same response always yields the same numbers. The model only writes the reading of a table it did not compute.

What is the difference between /v1/geo/answer-landscape and /v1/ai/visibility?

Answer Landscape collects the surfaces for you, including the engines. Visibility analyses answers you already have: you pass them in the answers array. If you are already storing engine output, the second is cheaper; if you are not, the first is the whole job.

Which engines are polled?

You choose them per call in the engines array, up to six. We do not publish a fixed roster on this page because it changes faster than a marketing page does — the API reference is the source of truth for what a request accepts today.

Can I measure a market other than my own?

Yes. gl and hl set the country and language of the search surfaces in the call, so the same query can be measured in several markets by making one call per market. Answer engines are polled per call as well, which means a market comparison is a set of calls you orchestrate rather than one request with a list of countries.

Do you publish benchmarks of how often each engine cites anyone?

We report what your own data says, not an industry average: POST /v1/geo/answer-landscape measures the overlap between organic results, the AI Overview, AI Mode and sampled engines for your queries, and the arithmetic is computed in code so the same response always yields the same numbers. If you quote a third-party figure, cite that source — the number you can defend in a meeting is the one measured on your own terms.

Keep reading

• Digital shelf monitoringThe same question for a product category, including the shopping surfaces.
• Dataswap vs SerpApiWhere a SERP-first vendor is the better fit, and where it is not.
• AI tools referenceVisibility, competitor briefs and the rest of the analysis routes.
• BlogNotes on how this is built and measured.