How the Index is made.

The question

When a buyer asks an AI assistant which product to use, which products does it name, and in what order? That is the whole measurement. We do not ask models to rate vendors. We ask the questions a buyer asks and record what comes back.

Prompts

Each category has a fixed set of 5 buyer prompts. They are written the way a founder or operator would type them: a stage, a team size, a constraint, a request for specific names. Prompts stay fixed between editions so movement reflects the models, not our wording. The exact prompts are printed on every category page.

Models

The same prompts go to 10 current-generation models, with web search enabled on every one:

• OpenAI: GPT-5.6 Sol (gpt-5.6-sol), GPT-5.6 Terra (gpt-5.6-terra), GPT-5.6 Luna (gpt-5.6-luna)
• Anthropic: Claude Opus 5 (claude-opus-5), Claude Sonnet 5 (claude-sonnet-5), Claude Fable 5 (claude-fable-5)
• Google: Gemini 3.6 Flash (gemini-3.6-flash), Gemini 3.5 Flash (gemini-3.5-flash)
• Perplexity: Sonar Pro (sonar-pro), Sonar Reasoning Pro (sonar-reasoning-pro)

Requests go through the DataForSEO LLM Responses API to each provider's own API, with the provider's web search tool turned on. This is the model with live web access in a clean session. It is close to, but not the same as, the consumer chat products: ChatGPT, Claude.ai and Gemini add memory, location and account history on top of the same models. We measure the reproducible layer. Each prompt is sent once per model per edition. Sampling is on, so one answer can vary; the count across 10 models and five prompts is the signal.

Answer share is computed across all 10 models. Category pages also show it per provider, so a vendor can see where OpenAI, Anthropic, Google and Perplexity disagree.

Scoring

Answer share is the number of answers that mention a vendor divided by the total answers for that category. Named first is the number of answers where that vendor appears before any other tracked vendor, divided by total answers. The model marks show whether each model named the vendor at least once.

A mention counts when the vendor's name or a known alias appears in the answer text. We maintain the vendor list per category and add any product that appears in the raw answers. If a product is missing, tell us and we add it for the next edition.

Sources

Where a model returns citations, we record the host of every cited URL and count them. ChatGPT via API returns no citations, so source counts under-represent it.

Limits

• Answers vary run to run. Small differences in share are noise. Large and repeated differences are signal.
• Alias matching is string-based. A common word used as a product name can over-count. We remove those cases when we see them.
• API models are not identical to consumer chat products. Consumer products add memory, location and personalisation.
• Prompts are in English and are not localised.

Cost and cadence

The 2026-09 edition covered 33 categories, 10 models and 1,650 answers, and cost $110.84 in API spend. Editions publish monthly. Raw answers are published in full on each category page.

Reuse

Charts, tables and figures from the Index are licensed CC BY 4.0. Reuse them in articles, decks and posts with the credit "Memetik Index" and a link to the category page. Every chart image carries the source in the frame and in its file metadata. Raw answers are model output and are published for verification.

Who makes this

Memetik is a Searchmaxxed publication. The Index is independent of client work. No vendor pays to be included or excluded.