Olson
How AI Answer Engines Name, Recommend and Cite a Brand
It measures how AI answer engines name, recommend and cite a brand: three figures kept apart, never a score.
- Figures per brand: named, recommended, cited
- 3
- Composite scores
- 0
- Report languages (Spanish or English)
- 2
The problem
A brand doesn't know whether, when someone asks about what it sells, an answer engine names it, recommends it or cites its site — or which sources sit behind what it says. And once those three are blended into a single number, you can no longer see which one is failing.
How it works
Olson puts a fixed bank of questions, written the way customers ask them, to answer engines through their APIs. Some state a need without naming any brand; others name the brand. Each question goes exactly as written, with no instructions and nothing about the company, so it measures what the engine says on its own. The questions are drafted without looking at the company's site. It queries OpenAI with web search, Gemini with Google Search, and Perplexity. It is named after Peggy Olson, from Mad Men. Draper helps us reach the customer; Olson helps the customer find us.
What it does
- 01
Questions the way customers ask
A fixed bank, with questions that name no brand and questions that do. They go exactly as written, with no instructions or company data, and are drafted without looking at the company's site.
- 02
Three figures, never a score
For each brand and its competitors it reports separately whether it is named, whether it is recommended and whether its site is cited, each rate with its counts. A model judges the recommendation and has to quote the sentence that shows it.
- 03
Sources and the site read
It shows which sources sit behind the answers. It never says what a cited page says or whom it names: when it fetches one, it records only what a reader without JavaScript received (status, words, links and structured data). It reads the brand's own site the same way, and checks which AI crawlers its robots.txt admits.
- 04
Report and second run
A PDF report, in Spanish or English, with one section per brand, a diagnosis and a focus plan. A second run is compared with the first question by question, with a range when the bank is large enough.
Deployment scope
There are two installations of the same code: Yáneken's, for its banners, and a personal one that runs as a Mac app, for other organisations. Each has several workspaces. Every run happens on command, never on a timer, because each one has a cost.
Limits
It measures what the engines answer through their APIs, not what each person sees in those engines' apps, which depends on their location, history and memory. It describes what the answers say, not what the pages say. A second run shows what changed between two measurements; it does not prove why.
Related projects
- ChispezaVoice AI for Retail Hiring and Sales Training
- AndreaAI Consumer Simulation for Marketing
- CifraAI Cash-Flow Intelligence for Multi-Entity Retail
Status and figures: September 2026
If you recognize this problem in your business, let’s talk. For an implementation with my team, the advisory work lives at Menlo & Oak.