Saylent

Answer surfaces

Google AI Overviews, Google AI Mode and Microsoft Copilot, read through your own SerpApi key and judged, scored and reported like any engine.

By default a run asks four model APIs: ChatGPT, Claude, Gemini and Perplexity. Many buyers never open those. They read the AI answer at the top of a Google results page, or the Copilot answer in Bing. Saylent calls these answer surfaces, and a run can add them to the engine set.

None of them has an official API. Saylent reads them through a SERP-data provider that you bring your own key for. The first provider is SerpApi. Surfaces are opt-in: a run that asks no surface is the same run it was before, and costs the same.

The three surfaces

SurfaceidWhat it readsSerpApi engine
Google AI Overviewsgoogle_ai_overviewThe AI Overview at the top of a Google results pagegoogle, then google_ai_overview when Google hands back a page token
Google AI Modegoogle_ai_modeGoogle's conversational AI Mode answergoogle_ai_mode
Microsoft CopilotcopilotThe Copilot answer in Bing searchbing_copilot

What each source returns, and which locale settings it honors:

SourceAnswer textCitationsCitations in answer orderCountryLanguageCity locationCan show no answer
The four model APIsyesyesyesnononono
Google AI Overviewsyesyesyesyesyesyesyes
Google AI Modeyesyesyesyesyesyesyes
Microsoft Copilotyesyesyesnononoyes

Set up the key

npx saylent keys set serpapi
npx saylent keys test

Get a key from serpapi.com/manage-api-key. The free plan has 250 searches a month. The key is SERPAPI_API_KEY, and it resolves like every other key: environment variable, then .env in the current folder, then ~/.saylent/config.json. It is never a flag and never a tool argument.

saylent keys test checks the key with SerpApi's Account API, which is free and does not count against your searches. It also prints how many searches your plan has left.

A surface you ask for without a key is refused by name, before anything is spent:

Google AI Overviews needs a SerpApi key (SERPAPI_API_KEY). Set one with
`saylent keys set serpapi` (free plan: 250 searches a month), or drop the
surface from the run.

It is never silently dropped from the run. --dry-run prints the same sentence on its Surfaces line and spends nothing.

Ask a surface

CLI

--surfaces adds surfaces to the engine set. --engines names the whole set and accepts surface ids too.

# the four model engines, plus two surfaces
npx saylent audit example.com --surfaces google_ai_overview,google_ai_mode --dry-run

# exactly these three sources
npx saylent audit example.com --engines chatgpt,claude,copilot

AUDIT_ENGINES and saylent.config engines accept the surface ids as well. An unknown name is refused, never dropped. The preflight prints one Surfaces line with the provider, the locale and the price per search:

  Surfaces  Google AI Overviews, Google AI Mode via SerpApi · country us · language en · priced at 2.5c/search (SerpApi Starter plan rate)

MCP

The audit tool takes the same three arguments: surfaces, country and location, and engines accepts the surface ids. A dry_run: true call reports keys_configured.serpapi (presence only) and a surfaces block with the surfaces requested, the locale, the provider and the refusal sentence when the key is missing. See The MCP server.

The app

The brand's run options page (/app/brand/<id>/questions) has an Answer surfaces group of checkboxes. When the deployment has no SerpApi key, the group is disabled and says why. Below it, Where the answer surfaces read from takes a Country, a Language and a Location. The estimate gets an Answer surfaces line priced per search.

The operator adds the key on Providers and models (/admin/providers), or as SERPAPI_API_KEY in the environment. The environment wins over a key saved in the console. Saving it from the console needs migration 0045_serpapi_provider.sql. The Test button uses the same free Account API.

The engine set is frozen like the question set, so adding or removing a surface for a brand starts a new baseline.

saylent.config.json

{
  "engines": ["chatgpt", "claude", "gemini", "perplexity", "google_ai_overview"],
  "surfaces": {
    "provider": "serpapi",
    "country": "us",
    "language": "en",
    "location": "Austin, Texas, United States"
  }
}

Every surfaces field is optional. They are only read when the engine set includes a surface. See Configuration.

Where the answer is read from

SettingWhere it comes fromSerpApi parameter
Country--country <cc>, else saylent.config surfaces.countrygl
Languagesaylent.config surfaces.language, else --locale (or config locale)hl
City location--location "<city>", else saylent.config surfaces.locationlocation

Country is a two-letter code (us, gb, de). Language is a BCP47 tag (en, pt-BR), sent lowercased. A location is capped at 120 characters. With none of them set, SerpApi answers from its default location.

--locale also translates the generated question set. To read the surfaces in one language without translating the questions, set surfaces.language in the config instead.

Microsoft Copilot takes no locale. SerpApi documents no country or language parameter for bing_copilot, so it answers from SerpApi's default location whatever you set. The preflight says so on the Surfaces line.

What it costs

SerpApi bills per successful search, by monthly plan (serpapi.com/pricing, checked 2026-11-02):

PlanPriceSearches a monthPer search
Free$0250-
Starter$251,0002.5 cents
Developer$755,0001.5 cents
Production$15015,0001.0 cent

Saylent prices every search at 2.5 cents, the Starter rate. That is the highest per-search price of any paid plan, so on a bigger plan the estimate runs high, never low. SerpApi does not count cached or failed searches. Its "Ludicrous Speed" plans bill twice per search, and the estimate ignores that.

Searches per answer:

SurfaceSearches per answerWhy
Google AI Overviews1 or 2 (the estimate uses 2)Google either puts the overview in the first results page, or hands back a page token that needs a second request to the google_ai_overview engine
Google AI Mode1
Microsoft Copilot1

A surface is sampled like any engine, so every draw is its own search. The judge call that reads each surface answer is billed like any other judge call. A question a surface did not answer (below) is still a search, so it is still billed, but it is never judged.

The searches go on your SerpApi bill, not your model-provider bill. The preflight estimate, the --max-usd check and the run's recorded cost all include them. More on the model-provider side: What it costs.

Rate limits

Each SerpApi plan has an hourly throughput cap: 50 searches an hour on the free plan, 200 on Starter, 1,000 on Developer. Over it, or out of monthly searches, SerpApi answers HTTP 429. Saylent retries twice with backoff, then records that answer as failed. A failed answer is an engine failure like any other: the report names the gap and the run exits 2.

When there is no AI answer

Google shows an AI Overview for some queries only. A surface can also come back with no answer blocks for a question. Saylent records that question, on that surface, as not shown:

  • it is a coverage gap, counted nowhere: not in the score, not in the band, not in share of voice
  • it is never an absence verdict, because there was no answer to be absent from
  • it is never an engine failure, and it does not change the exit code
  • the search was still made, so it is still billed

The live progress line counts them per surface, and the report marks each one "not shown" next to the question.

How a surface answer is judged and scored

Each surface answer becomes a normal answer row: the answer text, and its citations in the order the answer uses them. From there it goes through the same pipeline as an engine answer, so the judge, the score, the cited-page checks, Sources & opportunities and every report section treat it the same way.

The judge is the Anthropic family for all three surfaces. Google's two surfaces are Gemini-family answers and Copilot runs on OpenAI models, so Anthropic is the other family for each. With only one provider key, the run is single-family, as it is for any engine (Methodology).

Each surface asked gets its own per-engine band and counts in the overall score. The report labels it by name ("Google AI Overviews"), and the run records serpapi:google_ai_overview as the source that answered.

When a surface is asked, run.json carries run.surfaces: the provider and the locale the answers were read from. The run's call-cap backstop scales with the sources asked, 50 draws per source (350 with all seven), never below the base cap.

Verify

saylent verify asks the same engine set as the baseline, surfaces included, and reads them from the baseline's own locale (run.surfaces.locale), so the two runs are compared like for like. It needs the SerpApi key again. Without it, the verify is refused by name before anything is spent. The preflight prints the Surfaces line with "(the baseline's own locale)".

Add a provider

SerpApi is the only provider today. A second one is a small addition:

  1. Implement SerpProvider in packages/engine/src/adapters/serp.ts: one fetch per surface it serves, mapping its response into the shared answer block shape.
  2. Register it in SERP_PROVIDER_IMPLS (packages/engine/src/adapters/index.ts).
  3. Add its metadata to SERP_PROVIDERS in packages/engine/src/surfaces.ts: the env var, the surfaces it serves, its rate limit and its pricing page.
  4. Add its price per request to SERP_REQUEST_CENTS in packages/engine/src/rates.ts, with the date you checked it.

DataForSEO is a candidate. It is not implemented.

On this page