Study: agentic commerce, running shoes, Lebanon: AI knew the retailer. It knew the shoes. It rarely connected the two.
Mike Sport, a retailer in Lebanon, was recommended as a place to buy running shoes more often than any other retailer in the study: 52 times across 342 answers from three AI systems, against 48 for Decathlon. Across this test, 12 of 342 answers recommended a product Mike Sport sells and named Mike Sport as a place to buy it. Fewer than half of all answers, 149 of 342, named any shop.
The study at a glance
What was asked, of what, and when.
- Market
- Lebanon, adult running shoes
- Collected
- OpenAI and Perplexity on 1 September 2026, Gemini on 2 September 2026
- AI systems
- OpenAI, Perplexity and Gemini, with web search on and Lebanon as the location
- Scale
- 38 buying decisions, each asked three ways on three AI systems: 342 answers
How it was built
Four layers, each finished before the next began.
No layer could be shaped to fit the one after it.
- Market truth: 3,666 real demand observations, reduced to 105 specific buying situations, built without looking at the retailer's catalogue.
- Product truth: the retailer's live running range, 473 shoe models across 1,682 variants.
- What Mike Sport could win: which decisions its products could answer, established before any AI was queried. That was 26 of the 38 decisions tested.
- What AI answered: 342 recorded answers, from 38 decisions, three phrasings each, on three AI systems. They held 1,502 product recommendations, 612 merchant recommendations and 5,455 citations.
The retailer
Mike Sport was the most recommended place to buy.
| Place to buy | Recommendations |
|---|---|
| Mike Sport | 52 |
| Decathlon | 48 |
| Nike, selling direct | 22 |
| Running Warehouse | 20 |
| adidas, selling direct | 18 |
The lead over Decathlon is narrow. AI also described Mike Sport's stores accurately, down to the right outlet cities.
The shelf
What AI recommended, against what Mike Sport sells.
| What AI recommended | Count | Share of 1,502 |
|---|---|---|
| The exact product Mike Sport sells now | 348 | 23% |
| A product in its catalogue, but out of stock | 265 | 18% |
| A shoe family with no version named, where Mike Sport carries a current model | 270 | 18% |
| A specific version or variant Mike Sport does not stock | 93 | 6% |
| A product Mike Sport does not carry | 463 | 31% |
| A product we could not confidently match | 63 | 4% |
Shares are rounded. The 6% is the confirmed figure: 4% of recommendations could not be classified either way, mostly width requests the retailer does not publish data for.
The handoff
Where the store and the shelf met.
Most answers never tried to connect a shoe to a shop. The AI only named a shop when the buyer asked for one, and even then it rarely said which shoe to buy there.
| The buyer asked | Answers | Named a shoe | Named a shop |
|---|---|---|---|
| Which shoe should I get? | 189 | 84% | 16% |
| Where can I buy one? | 99 | 44% | 85% |
| Which shoe, and where can I get it? | 54 | 85% | 65% |
In 41 of the 342 answers a real handoff was on the table: the AI recommended a product Mike Sport sells and named somewhere to buy. Mike Sport won 12 and lost 9. Nine is a small number and we are not inflating it. In eight of the nine lost opportunities, AI never surfaced Mike Sport as a retailer. In one, Gemini surfaced Mike Sport and chose Decathlon instead.
The larger problem was being surfaced at the right moment, not being rejected after consideration.
Across AI systems
Which AI system you ask changes the answer.
Adding Gemini increased the corpus from 897 to 1,502 product recommendations, while the three principal product outcomes each moved by less than one percentage point.
| Outcome | OpenAI and Perplexity (of 897) | All three (of 1,502) |
|---|---|---|
| Sold now | 23.4% | 23.2% |
| Product not carried | 31.1% | 30.8% |
| Specific version or variant not stocked | 7.1% | 6.2% |
Run-level eligible-product capture is, among answers for decisions the retailer could fairly win, the share that recommended a product it sells. It ranged from 35.3% on Perplexity (18 of 51 runs) to 58.8% on Gemini (30 of 51), with OpenAI at 52.9% (27 of 51): a 23.5-point spread on identical prompts. It does not measure whether the answer named Mike Sport or any retailer.
Platform choice materially changes the measured result, so a single-platform audit is not a category read.
What this does not show
The limits of the study.
A finding is worth what its method can carry. These are the edges of this one.
- We recorded 342 answers. We did not estimate how often AI behaves this way in general.
- Three AI systems, one market, one category and one short window. OpenAI and Perplexity were collected on 1 September 2026 and Gemini on 2 September, and the first two were not run again.
- The runs were made through the systems' APIs with web search, and not in the consumer apps.
- The other 330 answers are not lost sales or failures. Many named no retailer at all, and Mike Sport did not have a fair claim to every buying decision.
- Every buying question was asked three different ways, because real customers do not all ask alike. That reduces reliance on one phrasing. It does not control for how sensitive the answers are to wording.
- Version and variant status was classified from how Mike Sport merchandises a product in its catalogue, and not from the manufacturer's complete global range.
- We observed what AI systems cited, named and recommended. We do not claim to know why any system chose what it chose.
- We did not estimate revenue. The study measures how decisions get answered, not how much they are worth.
- None of this says the retailer did something wrong. Mike Sport was recommended more often than any other shop in the study.
From the study
Your store and your shelf are two different assets in AI commerce.
Find what's worth changing next.
Bring the market, workflow or business question. We'll start there.