A “2 Hours Left” Timer Doesn't Work on AI: What to Check in Your Store

“2 hours left”, “only 3 left”, a crossed-out old price. Tricks like these rush a person, and stores have used them for decades. But more and more often the product is picked not by a person but by an AI assistant, such as ChatGPT with search. Does the same pressure work on it?

In May 2026, researchers writing in Harvard Business Review tried to answer this. Below: what exactly they tested, what they found and what the results do not say, plus what a store owner can check this week.

1. AI assistants already choose products

A person writes: “find me a work backpack under 2000 hryvnias” or “compare two coffee machines”. The assistant searches, compares product pages and recommends, and in some countries it can even complete the purchase. According to OpenAI’s documentation for sellers, a store gives ChatGPT a feed (a regularly refreshed CSV or JSON file) with identifiers, descriptions, prices, stock, images and delivery terms. Among the recommended fields the documentation names reviews, which can improve ranking.

A caveat right away. We do not know how many Ukrainians already buy through AI assistants: we found no Ukrainian study with a named methodology, and we do not quote figures floating around the web without a study name. OpenAI’s documentation does not say in which countries checkout inside ChatGPT works, and we could not confirm that it is available in Ukraine. So this is not a reason to panic but a reason to check what the assistant sees when it looks at your product page. That it may not buy by itself changes nothing: it can already recommend you or a competitor.

2. What Harvard Business Review tested

The article “Research: Traditional Marketing Doesn’t Work on AI Shopping Agents” was published in HBR on May 12, 2026. Authors: Jafar Sabbah (Bayes Business School, London) and Oguz A. Acar (King’s Business School, London).

What is known from the publisher’s page and from summaries:

• the researchers built a simulated storefront in which an AI assistant chose between products with different badges and promotions;
• they tested eight common marketing mechanics: scarcity (“only 3 left”), countdown, strike-through price, voucher, product bundle, rating and others;
• they used four AI models and four product categories; “thousands” of simulated choice rounds were run. The exact number is not given in the public description. The figure “16,000” that appears in some summaries was not confirmed by the original source, so we do not use it;
• according to summaries, the models were GPT-4.1-mini, GPT-5, Gemini 2.5 Pro and Gemini 2.5 Flash Lite. This comes from secondary sources: the full article is paywalled and we did not read it.

3. What worked and what did not

According to the HBR abstract:

• Rating (stars) was the only signal that consistently increased the chance that the assistant chose a product.
• Price consistently reduced choice: the higher the price, the lower the chance. In other words, a lower price helps.
• Scarcity, timers, strike-through prices, vouchers and bundles did not work reliably, and sometimes even reduced choice. The effect depended on the model and the category.
• More advanced “reasoning” models often looked more skeptical of overt pressure.

Here the sources differ slightly, and it is fair to say so. The HBR abstract names two stable factors: rating and price. One summary says only one of the eight mechanics behaved as it does with humans, and that is rating. Most likely the difference is that price is not a promotional mechanic but a property of the product itself. That is our assumption: the full text is not available and we could not check it.

The takeaway for a store is simple and matches common sense: an AI assistant is like a shopper who is in no hurry. A timer does not rush it; it looks at the price and at what other buyers write about the product.

4. What this study does not prove

Do not make more of it than it says.

• It is a simulation, not live shoppers and not real stores. How ChatGPT behaves in your niche with your products, the study does not show.
• Models change with every update. The authors themselves advise treating each model as a separate segment and regularly checking how it responds.
• The study does not say a timer hurts people. For people it may keep working: we did not test that. There is one conclusion: you cannot rely on a timer as your only argument, because part of your buyers already choose through an assistant.

5. What to check in your store

You do not need to buy or connect anything: everything below you can do yourself in a few hours.

1. The price on the website matches the price in the feed. Google Shopping (and ChatGPT, if you provide a feed) take the price from the file, and Googlebot compares it with the price on the page. According to Google Merchant Center help, the price in the page’s HTML must fully match the value uploaded to Merchant Center. If the price is loaded onto the page through JavaScript after it opens, Google may see a mismatch and the product may be disapproved. One common cause of a mismatch is a delay between updating the price on the site and in Merchant Center. If a sale has start and end dates, check the time zone. More about suspensions in the article “How to Avoid a Merchant Center Ban: 9 Checks”.
2. Reviews and rating are visible on the product page as text. The assistant reads the page as text. Reviews shown only by a widget that loads later may be invisible to it. If you are not sure, look at what appears in the page’s source code or ask your site developer. The rating must be honest: how a “perfect” five stars looks to people and Google we covered in “A 5.0 Rating Looks Fake”.
3. Stock and prices in the feed are current. If a product is sold out on the site but “in stock” in the feed, both the assistant and Google Shopping will show something untrue. The feed should refresh as often as prices change.
4. Your price is comparable to competitors. A lower price works consistently, but this is not a call to cut prices everywhere. See how price affects choice in product ads: “Price in Shopping Ads: Why People Click a Competitor”.
5. Look at yourself through the assistant’s eyes. Open ChatGPT with search and Gemini and ask them to pick a product in your category. Are you in the answer? What does it say about price and reviews? How to do this in 10 minutes we described in “What ChatGPT Says About Your Business”. And if an assistant invents what you do not have, read “How AI Hallucinates”.

6. Timer: remove it or not

We advise neither removing nor adding timers “because of a study”. If a sale is real and has an end date, the timer just shows it. It matters more that the date matches the data in your feed and is not an invented “2 more hours” every time. The study suggests something else: the core of a product page should be what reads the same to a person and to an assistant, which is an honest price, rating and reviews.

Summary

1. In May 2026 HBR published a simulation: an AI shopping assistant consistently responded to rating and price, while timers, scarcity, strike-through prices and vouchers worked unpredictably and sometimes hurt.
2. It is a simulation on four models and four categories, not live shoppers. How many Ukrainians already buy through AI assistants is unknown.
3. Check that the price on your site and in Google Shopping is the same and that reviews are visible on the page as text.
4. Keep your feed current and ask an assistant about your product once a month.

How we work with product advertising, see the sections on Merchant Center and Shopping.

Want a free audit of your Google Ads account? Send us a request: we will look at your account, compare prices and stock in your ads, in your feed and on your website, and tell you what is worth fixing.