We Asked Claude to Read Gymshark’s Black Friday Ads
Thirty-one and a half seconds after we asked Claude Code about Gymshark’s Black Friday ads, it had traced a sequence of offers: an early sale, changing discount codes, gifting messages, and a last-chance push. I was sur
Thirty-one and a half seconds after we asked Claude Code about Gymshark’s Black Friday ads, it had traced a sequence of offers: an early sale, changing discount codes, gifting messages, and a last-chance push.
I was surprised by both the speed and the pattern. All 30 ads it read offered up to 50 percent off, while the extra discount and the reason to buy changed.
Then we asked Cursor and Codex the same question. Both found the broad pattern, but they made different numbers of searches and requests for individual ads. Their workloads surprised me, too.
I build the vexqo API, which collected the public ad data in this experiment. The run was recorded on November 20, 2026. You can inspect the recorded MCP session, including the searches, timing, and credit costs.
Finding Gymshark’s 2025 ads
We gave Claude Code a specific task:
What did Gymshark run on Meta in the UK for Black Friday last year? Tell me the offers and the angles.
Claude had access to vexqo through the Model Context Protocol, or MCP, which lets an assistant call tools. Its first call searched for Gymshark and returned five ads, enough to identify the advertiser’s page. Its second call used that page ID, selected the UK, included active and stopped ads, and requested ads starting between November 1 and December 2, 2025.
The second search returned 30 ads out of 991 matches.
The searches took 4.5 seconds and 3.2 seconds. The full interaction, including the assistant’s work, took 31.5 seconds and used four vexqo credits. The agent’s own usage costs are separate. Those are measurements from this recorded run, rather than a promise about every future request.
How the discounts changed
All 30 ads in the sample advertised up to 50 percent off. The additional discount codes appeared at different points in the campaign.
The earliest sampled ads started on November 18, ten days before Black Friday. Some included an extra 10 percent discount with WELCOME10. A later ad offered new customers an extra 15 percent with NEW15. Ads with an extra 25 percent appeared from November 27; several named EXTRA25.
The messages shifted alongside the offers. Gifting appeared from November 22. Payday appeared from November 28. “Ends soon” appeared from November 30. The campaign timeline shows the dates and counts in the sample.
A team planning its own campaign could use those dates to decide what to test: when to introduce an offer, when to change it, and which message belongs at each stage.
It would still need to test those decisions with its own customers. The ads alone cannot establish whether the later discount improved conversion, whether gifting outperformed payday, or whether any of the offers made money.
What the sample leaves out
Claude read 30 of those matching ads. The documented sample favored ads with more impressions; it was not a random sample or a complete account of the campaign. Other offers, messages, and exceptions could appear in the remaining records.
Include the sample size when reusing the answer. If someone pastes the summary into a creative brief and drops the sample size, a finding about 30 ads can become a claim about everything Gymshark did.
I would also keep links to the original ads beside the summary. A start date is an observation. Calling an ad a successful evergreen creative is an inference that needs more evidence. The library’s multiple-versions label does not explain how the advertiser tested those versions or what the results were.
A separate Sephora example from our guide. Dates and creative details are visible; profitability and the reasons behind revisions are not. Source: Meta Ad Library, captured by vexqo on November 20, 2026.
If an angle does not appear in the first 30 ads, the next step is to search further before declaring that a competitor never uses it. Historical coverage also depends on what the library retains for the selected country.
The agent decides how much work to do
Cursor and Codex performed different amounts of research with the same question and tools.
| Agent | What it read in the recorded run | Time to answer |
|---|---|---|
| Claude Code | 2 searches | 31.5 seconds |
| Cursor command-line agent | 2 searches and 5 complete ads | 202 seconds |
| Codex | 17 searches and 9 complete ads | 571 seconds |
All three described the changing discounts, according to our session comparison. They also performed different amounts of work, so these timings cannot establish which model is generally faster or better at research.
A request for a quick campaign sketch needs a stopping point. A request for a thorough audit needs a coverage target. Without those instructions, the agent decides how many searches and detail requests to make.
For a first pass, I would try:
Find [brand]’s Meta ads in [country] for [date range]. Read up to 30 ads initially. Report how many you inspected and the total matches returned. Group the observed offers and messages by date, and cite supporting ad IDs. Separate observations from interpretations. Do not infer spend, sales, or profitability. Tell me what additional searches would most improve the answer before making more calls.
Writing the next brief
For the Gymshark example, a brief could propose keeping one sale headline and testing different supporting messages at defined stages. The team would measure the result in its own account. The public ads suggest a hypothesis to test.
The same approach can become a weekly practice. Collect comparable samples from the same competitors. Record what appeared, what changed, and what disappeared. Ask the assistant to help classify the evidence, then choose one question worth testing. Check the classifications and use code or a spreadsheet to verify counts before relying on them.
Collect ads each week, check the classifications and counts, and write a brief. Source: vexqo illustration.
For the next brief, I would ask for the pattern, the ads supporting it, and the evidence we would need before acting.
To try the same question, the MCP setup guide covers Claude Code, Codex, and Cursor. A vexqo trial includes 500 credits; the Claude session described here used four.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.
