SimplyCodes tested 35,398 promo codes provided by LLMs. 59% failed.

SimplyCodes asked ChatGPT, Gemini, Claude, and Perplexity for promo codes 9,337 times, then compared all 35,398 codes they returned with 12.5 million recorded promotions. We found that 59% of these codes failed, and failure rates were higher for the most frequently searched stores. Codes from the 50 most searched stores worked 24% of the time, versus 56% from the 50 least-searched.

We set out to learn if LLMs answer shopper queries with accurate promo codes.

The 2026 SimplyCodes AI Promo Code Study asked Claude, ChatGPT, Gemini, and Perplexity, the AI assistants built on large language models (LLMs), 468 pre-registered questions about promo codes at 150 stores, with web search on, sampling each question three times (five for Gemini). The assistants returned 35,398 codes. Each one was matched against the SimplyCodes Code Truth Store, a database of 12.5 million recorded promotions, and classified as working, dead, wrong merchant, never seen, held but never tested, or offered for a retailer other than the one asked about, using the same verdict logic SimplyCodes applies to every listed code.

More than half of the codes in AI answers are unusable.

Of the codes we could settle against the SimplyCodes database, 59% failed. These are codes an assistant produced, not codes SimplyCodes listed and tested, so the rate is not comparable to our own verification figures. We arrived at that figure after a pre-registered hand audit, described in "How we verified our results," found that some of the codes that had not been captured in our database were real. Based on those results, we corrected our initial figure of 67% downward to 59%.

AI assistants offer more failed codes for the most frequently searched for stores.

The SimplyCodes data strongly suggest that the more shoppers search for a store's codes, the worse the AI answers get. We split the 150 stores into three bands of 50 by how often shoppers search for their codes. Each band contained roughly 12,000 returned codes.

At the 50 most-searched stores, roughly one code in four works (24%). The middle band works 44% of the time. At the 50 least-searched, 56% work.

This is likely because AI assistants are pulling inaccurate information from their web searches. For the biggest, most heavily searched brands, hundreds of pages are competing to rank for their codes, and most of them are stale, recycled, or invented for traffic. We laid out how that pollution feeds back into the models before we had numbers for it. Online information on codes published by smaller stores is often more accurate because there are fewer marketing pages trying to rank for them.

ChatGPT provided the most working codes per answer. Claude, which provided significantly fewer codes, was the most accurate.

SimplyCodes analyzed AI assistant responses to requests for promo codes in two ways. First, we looked at how accurate the provided codes were. And then we counted how many codes each model provided. The two numbers together tell us how many usable codes each answer is likely to contain.

Claude offered 2,883 codes and 47% worked (2.2 codes per answer, 0.8 working). Gemini offered 9,182 at 43% (4.1 per answer, 1.3 working). ChatGPT offered 10,753 at 41% (8.0 per answer, 2.4 working). And Perplexity offered 12,580 at 37% (9.3 per answer, 2.2 working).

Claude provided the fewest codes, but also the highest share of working ones, netting out to less than one working code per answer. ChatGPT had the second-worst accuracy, but offered so many codes per answer that a shopper willing to test them all ends up with the most working codes of any assistant.

AI assistants often "clipped" or truncated promo codes.

The study's signature discovery is a behavior that has not been previously documented, namely that assistants often remove or "clip" the first few characters from promo codes. For example, WELCOME50 comes back as OME50, FREESHIP-25 becomes IP-25, SAVE20 becomes AVE20, etc. We also observed that assistants will often return both the whole code and its clipped twin in one reply. One ChatGPT answer about a single apparel brand contained four codes in a row, each missing its opening characters. The pre-registered audit confirmed clipped codes against merchants' own published codes, character for character.

Across the study, SimplyCodes found 539 full-and-clipped pairs across 109 of the 150 stores. Chance does not explain that. Randomly reshuffling the same codes between stores 200 times never produced more than a few dozen pairs, against the 539 we recorded.

We do not have a clear model for how this happens. Language models are probabilistic systems being asked for an exact string, which makes character-level fidelity a plausible place to look. One possible explanation is that coupon sites hide part of a code until a visitor clicks, so the model only "sees" and transcribes the visible tail. But a masked code appears in the raw text of fewer than one percent of answers, far too rarely to account for the clipping we measured, so the mechanism remains a hypothesis.

AI assistants will fabricate promo codes.

SimplyCodes invented six store names, and conducted a search to ensure we didn't inadvertently duplicate a real business' name. Because these stores are fake, there are no valid promo codes for them. When we asked AI assistants to provide promo codes for these fake stores, three out of four obliged. Claude answered 54 times and invented nothing. ChatGPT invented 156 codes, Gemini 175, Perplexity 301.

We also asked the AI assistants to tell us about the invented stores, and they often replied with an answer describing a real company with a similar name. ChatGPT did it 50% of the time, with Gemini at 64%, Perplexity at 63%, and Claude at 4%.

This finding only applies to questions submitted to AI assistants through an API. When we repeated our 20 questions including fake stores in the consumer ChatGPT and Gemini apps, both refused to provide codes for the fabricated stores and provided roughly a quarter as many codes for the real stores. This means fabrication is a bigger concern for shoppers accessing codes via third-party tools powered by an API than for shoppers using AI assistants' native apps. What a verification layer for those tools has to do is a separate question we have written about.

How we verified our results

Our original analysis compared the promo codes we captured from AI assistant answers with the 12.5 million recorded promotions in the SimplyCodes database. We were able to match 81% of generated codes with entries in our database, leaving 19% of codes unaccounted for. To find out if these codes were hallucinations or simply represented coverage gaps in our database, we ran a manual validation step.

We started by drawing a pre-registered sample of 100 of the "missing" codes and hand-checking each one against merchant websites and the broader internet, under rules and thresholds written down before the exercise began.

We found:

• 35 real codes our database had never collected; only 23 of these would still help a shopper
• 15 real codes the assistant had clipped or mangled
• 33 with no public trace anywhere, reported as unverifiable
• 3 unable to assess
• 14 records that were not promo codes

Based on this finding, we adjusted our headline failure rate to 59% from an initial 67% that had assumed all codes missing from our database were either fictitious or unusable.

The most conservative interpretation

A code lacking an intelligible web trace is not necessarily a fabrication, because codes die off with age, are unlisted, or may be generated for an individual customer; we report those as unverifiable rather than fabricated. Also, as noted above, fourteen strings were never codes at all. Two were usernames lifted from coupon-site pages, and one was SIMPLYCODESD, our own brand name, which an assistant displayed as a merchant's discount code.

But even if we assume every code unmatched to our database is potentially usable, 41% of the codes we could settle against our database were provably dead or belonged to another store. That floor rests on no assumptions.

What we recommend for shoppers

• Treat promo codes in AI answers as potentially usable codes and plan on checking them all. Here is how to test one without wasting an order.
• Manage expectations when seeking codes from popular brands, since we found three in four codes appearing in answers to queries about the most popular stores did not work.
• Use a verification service like SimplyCodes to better assess the likelihood of a code working. (Every code is tested before it is listed, and each one shows its testing state and freshness, so you can see when it was last confirmed.)
• Remember that AI assistants will truncate codes, so a code starting with AVE or ELCOME may be worth trying with the most likely first letters filled in.

Frequently asked questions

Do promo codes from ChatGPT work?

In the 2026 SimplyCodes AI Promo Code Study, 41% of the ChatGPT codes that could be settled against the SimplyCodes database worked. ChatGPT returned 8.0 codes per answer, second only to Perplexity at 9.3, and that yielded about 2.4 working codes per answer — the highest of the four assistants tested.

Which AI assistant gives the most accurate promo codes?

Claude. In the SimplyCodes study, 47% of codes captured from its answers worked. But Claude also returned the fewest codes, just 2.2 per answer, so shoppers end up with fewer working codes than they would if they used a less accurate model providing a lot more codes.

Why do AI promo codes fail most at big stores?

Heavily searched brands attract hundreds of competing coupon pages, many of which are stale or invented for traffic. AI assistants typically repeat what they find on the internet if they're using web search tools. The SimplyCodes study found that codes for the 50 most-searched stores worked 24% of the time versus 56% at the 50 least-searched.

Does AI make up promo codes for stores that don't exist?

Through their APIs, yes. In the SimplyCodes study, AI assistants were asked to provide codes from six invented stores. In response, ChatGPT, Gemini, and Perplexity produced 632 codes for businesses that do not exist, and Claude produced none. The consumer ChatGPT and Gemini apps refused the same questions, so this finding describes the API, not the app.

Methodology

The 2026 SimplyCodes AI Promo Code Study asked 468 pre-registered questions across 150 stores in three demand bands, plus six invented control stores, to Claude, ChatGPT, Gemini, and Perplexity with web search on, three samples per question (five for Gemini), in August 2026.

The question battery was locked and hashed before collection, and no question was added or reworded afterwards. Codes were extracted by a language model under an anti-invention guard and matched against the SimplyCodes Code Truth Store of 12.5 million promotions. Codes were assigned one of these categories: working, dead, wrong merchant, never seen, held but never tested, or offered for a retailer other than the one asked about.

The never-seen audit was pre-registered with its decision thresholds written before validation, and its mid-audit amendment adding the clipped-code class is part of the public record. A clipped code is defined as a returned code that is a suffix of another code returned for the same store, where the suffix is at least four characters and at most four leading characters were removed. 9,337 answers, 35,398 codes, collection cost $232. The measurement was designed and run by Sean Fisher.

Limitations and disclaimers

• The volume and fabrication findings describe the API, not the consumer app.
• We adjusted our headline failure rate based on the hand-validation step described above.
• "No public trace" is reported as unverifiable, not as a confirmed fabrication.

Citation

SimplyCodes Research. (2026). SimplyCodes tested 35,398 promo codes provided by LLMs. 59% failed. SimplyCodes. https://simplycodes.com/blog/ai-promo-codes-study