Comparisons
Manual prompt checks vs RankAI: what changes
RankAI changes three steps in a manual prompt check: how answers are collected, how separate observations stay separate, and how each one is verified before approval.
If you are comparing a manual prompt check with RankAI, three steps change: how answers are collected, how separate observations are kept apart, and how each observation is verified before it is published. In short: collection per engine, separation of observations, and verification before approval. In a manual check, the person writing the prompt opens the engines one by one and notes what they see. With RankAI, collection runs through the public pages that name ChatGPT, Perplexity, Gemini, Grok and Google AI. The third step, verifiability, carries the most editorial weight: an observation with no verification criterion does not make it into the version to be approved.
Side by side on a concrete case
Take a prompt about AI visibility and an editorial plan. In the manual check, the person writes the prompt, reads the answer on one engine, copies it into a document and adds a comment. With RankAI, the same observation is collected per engine, kept distinct from the others and put through the three verification fields before approval. The question is the same; what differs is what stays traceable afterwards.
The three steps, each with a verifiable criterion
- Collection per engine. Criterion: every answer is tied to the engine that produced it. Gemini and Google AI Overviews are not the same observation and must be kept separate. Claude and Grok are measured only after you connect them.
- Separation of observations. Criterion: an observation that merges two engines into a single sentence is not verifiable and must be discarded or split. An API response is not a person's screen, and should not be treated as one.
- Verifiability before approval. Criterion: an observation passes only if it records the engine, the prompt and the collection date. If one of the three fields is missing, it stays out of the version to be approved.
A filled-in record: what passes and what does not
An observation that passes all three fields
Engine: Perplexity. Prompt: “how do you build an editorial plan for AI visibility”. Date: the day of collection. Result: the observation goes into the version to be approved, because all three fields are present and the engine is named.
An observation missing one field
Engine: Gemini and Google AI Overviews listed together. Prompt: present. Date: present. Result: the observation does not pass, because two distinct observations are fused into one and the engine field is no longer verifiable. It has to be split before it can go back in.
Automatic publishing comes later
The default in RankAI is human approval of a version. Automatic publishing is turned on per site and still goes through the checks: it does not replace the three steps, it assumes them. So the detail on automatic publishing does not change the answer to the opening question, but it does clarify what happens after approval.
This is RankAI's working proposal, not a universal standard: the three-field record (engine, prompt, date) is meant to make comparable what a manual check tends to merge into a single impression. That last point is a qualitative observation, not a measured data point. The method we propose is simple: before approving a version, fill in a record for each observation; if a field is left empty, do not patch the observation from memory, go back to collection. That way verifiability is not a final judgment but a step with an explicit stop rule. The filled-in record above shows the most common edge case: two engines named together, which look like one observation but are not. Keeping the record separate per engine also makes the editorial plan easier to read later on, because every line points to a specific collection rather than a memory.
Verifiable limit: RankAI does not measure 豆包, 元宝, Kimi or 千问, and Claude and Grok appear only after you connect them. A full comparison with a manual check that includes those engines is therefore not possible using RankAI data alone. For the complete picture of steps and checks, see the feature page.