SimplyRank Research

Who runs it

The research programme is run by Karim Nassar, the founder of SimplyRank. He is accountable for the methodology set out on this page and for every figure published under the SimplyRank Research byline.

SimplyRank is a small, founder-operated company. The research is not outsourced to an agency and there is no anonymous panel of analysts behind the name: when a number appears here, one named person is answerable for it. Methodology questions and corrections go to [email protected].

What each scan measures

A scan produces the same set of measures for every brand tracked. These are the definitions used inside the product, and they are the definitions behind any figure we publish.

• Visibility — the share of a scan’s prompts where the brand was mentioned in the answer — prompts with a mention, divided by the prompts that ran, expressed as a percentage. Nothing is weighted, capped or adjusted after the fact.
• Position — where the brand sits in an answer when it is mentioned, averaged across mentions. Lower is better: an average of one means the assistant named it first every time it appeared.
• Sentiment — each mention is classified positive, neutral or negative, then aggregated into a percentage split.
• Citations — the external sources an assistant names or links when it answers. Perplexity cites on every answer; the other three cite inconsistently, which is itself a finding.
• Gap analysis — the prompts where a tracked competitor is named and the brand is not.

The assistants we query, and their versions

Every prompt goes to four assistants independently, because the same question routinely gets different answers from each. The versions in production on the date at the foot of this page are:

• ChatGPT — OpenAI, gpt-4o.
• Claude — Anthropic, claude-sonnet-5.
• Perplexity — sonar-pro.
• Gemini — Google, gemini-2.5-flash.

How a scan runs

Prompts are the questions a customer configures. They are written to look like what buyers actually ask an assistant — "what is the best CRM for a five-person agency?" — rather than brand-name lookups, because the point is to observe discovery rather than recall.

Each active prompt is put to all four assistants on a schedule set by the plan: weekly on the entry plan, daily or weekly on Starter, and daily, weekly or monthly on Pro and Advanced. Scans can also be triggered manually within plan limits.

A scan is therefore the full cross-product of one brand’s prompts and four assistants: every active prompt is put to each assistant once, so a scan returns four answers per prompt and nothing is sampled or skipped. Sample sizes are per customer rather than a shared corpus — we do not run a fixed public panel of prompts, so any figure we publish states the prompt set behind it.

Model versions change when providers retire them. Claude scans ran on Sonnet 4 until Anthropic retired that model in June 2026 and moved to Sonnet 5 the same month. Every stored response carries the model version that produced it, so older figures can still be read against the model that generated them.

What we exclude

Published analysis only ever reads completed scans: sessions that recorded results for at least 98% of their prompts. Failed and abandoned sessions are excluded from every score, chart and published figure. One caveat we state rather than hide: when a single provider fails terminally mid-scan, the session can complete on the remaining assistants and stay in the dataset — those sessions represent fewer than four models, and per-model charts reflect whichever assistants actually answered. Repeat scans of the same brand on the same day are deduplicated to a single canonical result, so a manual re-run cannot double-weight a day.

What this data cannot tell you

• Not attribution — AI answers do not reliably expose outbound clicks, so visibility is an upstream discovery signal, not a revenue model.
• Not every assistant — Copilot, open-source models and region-specific engines sit outside the four we query.
• Not a single-scan read — answers vary between runs, so one scan’s movement is not a trend. Comparisons should be made over weeks.
• Not a census — results describe the prompts configured for a brand, and nothing beyond them.