3LA Benchmark EngineAI visibility measured across an industry, not one answer at a time
Who does AI recommend in your industry, and how stable is it?
The 3LA Benchmark Engine asks ChatGPT, Gemini and Perplexity the same buying questions your customers ask, and repeats them over several days. You see who gets recommended, how often, which sources the answers rely on, and which differences are real rather than random variation.
What is the 3LA Benchmark Engine?
The 3LA Benchmark Engine is a measurement engine for AI visibility. It compares the leading businesses in an industry on the same questions, the same AI assistants and at the same time. Every answer is stored with the sources the assistant cited, then read for which businesses it mentions. Because the same setup runs repeatedly, every share comes with a confidence interval, and stable positions can be told apart from random swings. The result is an industry view of who the AI assistants actually recommend, how often, for which kinds of question, and which websites those recommendations rest on. The engine is part of the Three-Layer Approach framework, where the AI layer sits alongside human readability and search.
What did the first industry study show?
The first study measured online retail in the United Kingdom and Norway in October 2026: 1,728 AI answers from four runs over three days. The top of the lists was very stable from day to day, but the three AI assistants often disagreed with each other about who is best. In other words, which assistant a customer uses matters more than which day they ask. The findings are preliminary, and retailer names are not published.
1 of 6
UK categories where all three search-enabled assistants agreed on the number one retailer. In Norway it was 6 of 12.
10 of 10
of ChatGPT's most-mentioned UK retailers were the same with and without web search (7 of 10 in Norway). The leaders live in the model's memory.
1 in 5
UK retailers outside the top were mentioned on only one of the days (about 1 in 4 in Norway). The top is stable; the tail shifts.
38%
of the sources AI cited in the UK were the retailers' own sites. The largest third-party source was the consumer body Which?.
Stable over time, divided among themselves
Share of categories, preliminary figures from four runs
Which sources do AI answers rely on?
Share of all cited sources, selected source types
Why is one AI answer a day not enough?
AI assistants do not answer the same way every time, even to exactly the same question. A dashboard that collects one answer per question per day therefore shows a lot of random variation as if it were real change. In our UK data (576 answers), two answers taken minutes apart differed as much as answers taken on different days:
| Assistant | Same run | Different days |
|---|---|---|
| Gemini (with Google Search) | 0.76 | 0.76 |
| ChatGPT (memory only) | 0.75 | 0.75 |
| ChatGPT (web search) | 0.70 | 0.70 |
| Perplexity | 0.62 | 0.57 |
Jaccard overlap between the sets of retailers mentioned (1 = identical). For Gemini and ChatGPT, day-to-day movement was indistinguishable from sampling noise; Perplexity showed a small amount of real drift. The Benchmark Engine is built to separate signal from noise.
Known uncertainty
Every question runs several times with the same setup. All shares are reported with a Wilson confidence interval, so small differences are never sold as big changes.
Memory versus search
A separate measurement arm asks ChatGPT the questions without web search. You see whether your business is a brand the model already knows, or only something it finds when it searches.
Industries and markets side by side
A fixed question set per industry makes businesses directly comparable. The same method runs in English and Norwegian, in the UK and in Norway, so differences between markets become visible.
How does the measurement work?
The measurement uses the AI providers' own interfaces with web search switched on: Gemini with Google Search, Perplexity and ChatGPT with web search. Every answer is stored with timestamp, assistant and the sources it cited, then read for which businesses it mentions. Google AI Mode, Copilot and Claude are not yet part of the benchmark.
01
Industry universe
We map the businesses to be measured, with domains, alternative names and former names.
02
Real customer questions
The questions cover advice, comparison and purchase, written the way customers ask in their own language.
03
Four measurement arms
Gemini, Perplexity and ChatGPT with search, plus ChatGPT without search to measure memory.
04
Repeated runs
The same setup runs over several days, so we can see which positions hold.
05
Analysis
Answers are read for businesses and sources, and every figure gets a confidence interval before it is reported.
Read more about how 3LA measures in How 3LA works and the GEO guide.
What questions does it answer?
Every benchmark produces the same measurements, so results can be compared across businesses, AI assistants and points in time.
| Measurement | The question it answers |
|---|---|
| Mention rate | How often are we recommended, per AI assistant, category and language, compared with competitors? |
| Change over time | Is our position stable, or do we come and go from run to run? |
| Source anatomy | Do AI answers rely on our own pages, consumer tests, price comparison, reviews or forums? |
| Own-domain citation | Does the assistant link to our own website, or only mention our name? |
| Memory versus search | Does the model know us without searching, or do we only appear when it searches? |
| Answers without businesses | Which questions get answers that name no business at all? Those are open positions. |
| Unexpected players | Who do the AI assistants mention that we did not count as competitors? |
What data do you receive?
Every benchmark is delivered as a PDF report, a CSV file with every measurement and raw JSON data for each run. The raw data contains every individual AI answer with the assistant, timestamp, the businesses mentioned and the sources it cited. That lets you build your own analysis, connect the data to BI tools or share it with clients.
- PDF and CSV for the whole benchmark and for each run
- JSON with every AI answer and its sources, for each run
- API and MCP: the 3LA platform has both today. Access to benchmark data via API key and dedicated MCP tools is in development.
Illustration of the format, one AI answer
[
{
"edition": "vol_retail_GB_2026Q4",
"sampled_at": "2026-10-11T05:12:07Z",
"provider": "openai_web",
"language": "en",
"prompt_id": "en_elec_comm",
"mentions": ["retailer-a", "retailer-b", "retailer-c"],
"citations": [
{ "domain": "retailer-a.co.uk", "class": "own" },
{ "domain": "which.co.uk", "class": "consumer_test" }
],
"parametric_only": false
}
]
Which plan is right for you?
You can start self-service with Brand Analyzer, commission a benchmark for your whole industry, or deliver the analysis under your own name as an agency. All subscriptions are on the pricing page.
For brands
Brand Analyzer
from €199 /month, Team
- Your own questions and competitors
- ChatGPT, Claude, Gemini, Perplexity and Mistral
- Sentiment, share of voice and alerts
Project
Industry or competitor benchmark
Priced by scope
- Your position against 10–20 competitors, or the whole industry
- Repeated runs with confidence intervals
- Report, CSV and JSON with every answer
For agencies
Agency
€499 /month
- White label: deliver the analysis under your own name
- Multiple clients in one setup
- 5,000 credits per month
What do users say about 3LA?
The reviews below are about the 3LA platform, which the Benchmark Engine is part of.
I highly recommend the 3LA platform! It only took a short time from when I joined the waiting list until I was approved and could receive three free analyses. My first impression was brilliant. The platform provides genuinely valuable insight.
The Three-Layer Approach (3LA) is a groundbreaking service! In the report, you don't just learn what is wrong with your site, but also what you should do to improve it – specifically tailored to you.
I was lucky enough to test 3LA in the beta version and found it to be a very useful and user-friendly tool – regardless of technical expertise. The analysis you receive provides concrete, actionable suggestions for website improvements.
Frequently asked questions
Is the measurement the same as what a customer sees in ChatGPT?
Not exactly. We measure through the AI providers' own APIs with web search switched on, while signed-in users may get answers shaped by their history. That is why we ask the same questions many times and report shares with confidence intervals, rather than presenting a single answer as the truth.
How often is a benchmark run?
A benchmark runs on several consecutive days with the same setup, so we can separate findings that hold from variation that comes and goes. Ongoing monitoring after the first benchmark is agreed per engagement, for example monthly or quarterly.
Which AI assistants do you measure?
The Benchmark Engine measures Gemini with Google Search, Perplexity and ChatGPT with web search, plus ChatGPT without search to see what the model remembers. Google AI Mode, Copilot and Claude are not yet included in the benchmark. Brand Analyzer, which customers run themselves, also covers Claude and Mistral.
What data do we receive?
You receive a PDF report, a CSV file with every measurement and raw JSON data for each run. The raw data contains every individual AI answer with the assistant, timestamp, the businesses mentioned and the sources the answer cited, so you can run your own analysis.
Can we access the data through an API or MCP?
The 3LA platform already has both an API and an MCP server, covering audits and brand data among other things. Access to benchmark data via API key and dedicated MCP tools is in development. Until then, benchmark data is delivered as files.
Who is behind the 3LA Benchmark Engine?
3LA is developed by Nettpilot AS, a Norwegian company (org. no. 934 178 580). Nettpilot sells AI visibility consulting. For that reason Nettpilot is measured in its own studies but never ranked, and the method is never adjusted in anyone's favour.
See how AI recommends your industry
Tell us which industry and which country you are interested in. Within two working days we will reply with a proposed question set, a list of businesses that should be measured, and a price for the project.
Would you rather try it yourself first? Start with a free AI check of your brand.
Thank you, we have received your request.
You will hear from us within two working days with a proposed setup and price.