Research methodology

Dated evidence—not a universal ranking.

Every useful claim should trace to a buyer question, system, exact model, check date, returned answer, competitor observation, source note, and reviewed action.

Models vary by plan. Saved answers keep their model and date.

Request a free review

What the benchmark measures

Insights submits buyer questions to named model APIs through OpenRouter. The application retrieves search results through Exa and supplies that evidence with the question. Models return a structured recommendation list of up to eight entries.

One search result set is shared across the repeated answers for a question in a run. Those answers test variation under shared retrieval; they are not independent web searches.

This is an API benchmark, not a recording of answers in ChatGPT, Claude, Gemini, Grok, Meta AI, or DeepSeek consumer apps. App-specific search, instructions, location, account history, and personalization can produce different answers. Consumer-app spot checks must be labeled separately.

Controlled check

Keep the question stable and the evidence attributable.

Choose a real question

Use natural category, comparison, trust, alternative, or purchase-intent language.

Record context

Preserve brand, category, location when relevant, competitors, and the exact submitted question.

Run named models

Use the active registry and never silently substitute a model after failure.

Preserve evidence

Store exact model identifier, display name, timestamp, raw response where policy permits, and parsed findings separately.

Review findings

Separate mentions, recommendations, order when meaningful, competitors, citations, evidence gaps, and uncertainty.

Create one action

Tie the strongest supported gap to a URL or asset, rationale, owner, effort, and recheck timing.

Google, Gemini, and AI Overviews

How we keep these related experiences clear.

Google’s AI search experiences, including AI Mode and AI Overviews, are connected to Gemini-powered AI systems. In The Answer reports a Google/Gemini check separately from other systems so businesses can understand how their brand appears in an important Google AI context.

This is not the same as claiming every live Google Search AI Overview will match the report. AI Overviews can vary by query wording, location, account state, personalization, rollout status, and timing.

For that reason, every report labels the system, model, date, buyer question, and evidence used in the check.

We separate

Google/Gemini context
Live Google Search AI Overview behavior
Other model checks such as ChatGPT, Claude, DeepSeek, Grok, and Muse Spark

The goal

Show whether your brand appears in important AI answer contexts and what evidence may be missing—without claiming control over Google’s live answers.

Current model registry

Consumer surface and direct check stay distinct.

Registry updated 2026-09-27. The account dashboard lists the active model versions available to each plan.

Consumer surface	Current In The Answer model	API source evidence	Search context	Failure handling
ChatGPT	GPT-6 Astra	References to supplied sources, when returned	Application-provided search results; native web browsing not enabled in this benchmark	Show unavailable; do not substitute
Gemini and Google AI search experiences	Gemini 3.8 Flash	References to supplied sources, when returned	Application-provided search results; native web browsing not enabled in this benchmark	Show unavailable; do not substitute
Claude	Claude Opus 5.5	References to supplied sources, when returned	Application-provided search results; native web browsing not enabled in this benchmark	Show unavailable; do not substitute
DeepSeek app (separate experience)	DeepSeek V4.1 Flash	References to supplied sources, when returned	Application-provided search results; native web browsing not enabled in this benchmark	Show unavailable; do not substitute
Grok on X, web, and mobile	Grok 4.7	References to supplied sources, when returned	Application-provided search results; native web browsing not enabled in this benchmark	Show unavailable; do not substitute
Meta AI (Thinking mode)	Muse Spark 1.3	References to supplied sources, when returned	Application-provided search results; native web browsing not enabled in this benchmark	Show unavailable; do not substitute

Historical reports preserve the exact checked model. Google AI Overviews and Meta AI are consumer experiences; Gemini 3.8 Flash and Muse Spark 1.3 are named direct routes in the current lineup.

What we can observe

Directly supported fields.

Whether the brand appears
Recommendation or mention language
Ordering when meaningfully measurable
Competitors returned
Visible citations or sources
Model and check date
Provider failure or partial completion

What we infer cautiously

Directional—not causal.

Potential category clarity gaps
Potential proof or comparison gaps
Potential source and citation gaps
Potential entity-information gaps
Which action is most useful to test first

Limitations

AI answers vary by model, version, date, settings, available information, and user context. One question cannot represent an entire market. Repeated appearances are not market share. A before-and-after change does not by itself prove that one page edit caused the result. Sources are recorded only when returned in the evidence. Retrieval of a page does not by itself mean the answer cited or recommended it. Recommendation rate uses valid completed answers as its denominator; failed and skipped checks remain visible in collection coverage. Position is measured within the returned list, not as a universal search ranking. Model, prompt, search, or identity-rule changes start a new comparison segment.

AI tools and sources

The full model details.

OpenAI

ChatGPT

ChatGPT · Current model: GPT-6 Astra

44%of U.S. adults report using ChatGPT

Used directly for product research, comparisons, recommendations, and work.

U.S. usage source

Google

Gemini + Google Search

Gemini and Google AI search experiences · Current model: Gemini 3.8 Flash

24%of U.S. adults report using Gemini

The direct Gemini check can be reviewed alongside Google-visible evidence. It does not reproduce AI Overviews or AI Mode.

Google reported 1.5B monthly AI Overview users in Q1 2025; the In The Answer run checks Gemini, not a live Google AI Overview.

U.S. usage source Related reach source

Anthropic

Claude

Claude · Current model: Claude Opus 5.5

6%of U.S. adults report using Claude

Used for detailed research, writing, coding, long documents, and careful comparison work.

U.S. usage source

DeepSeek

DeepSeek app (separate experience) · Current model: DeepSeek V4.1 Flash

Not measuredNo comparable U.S. adoption percentage verified for this update

An API check using web evidence supplied by In The Answer. It does not reproduce the DeepSeek app or its native browsing.

No comparable U.S. usage source verified

xAI

Grok

Grok on X, web, and mobile · Current model: Grok 4.7

8%of U.S. adults report using Grok

Used for current-event, public-conversation, research, and recommendation questions.

U.S. usage source

Meta AI

Meta AI (Thinking mode) · Current model: Muse Spark 1.3

14%of U.S. adults report using Meta AI

This is a direct Muse Spark 1.3 API check. The Meta AI app can add different context and tools, so its answers may differ.

The Meta AI consumer product can apply different tools, context, and instructions than this direct model route.

U.S. usage source Related Meta AI source

Publication safeguards

Research is published only when the dataset can support it.

Dataset gate

Minimum sample thresholds
Complete underlying rows
Documented inclusion rules
Documented date range

Privacy gate

Anonymization
No customer-identifying information without permission
Retention and deletion review

Integrity gate

Human review
No invented findings
No publication when data is too small or biased
Methods and caveats published with findings

Learn more

See how we check. Choose your next step.

This page was checked on 2026-09-26. Read how we check. Or choose a page below.