What AI shopping agents will pay for, and what they ignore

Your next customer might not be a person. It might be software shopping on someone's behalf.

That raises a fair question. Do your offers work on a machine?

New research gives us a first look. The answer is useful. Clear, countable offers land. Clever persuasion mostly does not. (Countdown timers may want a quiet word.)

So it is worth knowing which of your promotions speak to a machine, and which only speak to a mood.

What the research found

Utpal Dholakia ran five experiments with 29,070 responses across 13 AI models. The models played the shopper. They put a price on different parts of an offer.

The free shipping result stands out. 10 of 17 model configurations valued free shipping at exactly $6.99.

The same figure, again and again. Human shoppers rarely agree that neatly.

Now compare a softer cue. A 4.7-star rating, compared with a 4.4-star rating, was worth $14 to one model and $47 to another. That is a wide gap for a 0.3-star difference.

The models disagreed with each other. That matters. If you lean on star ratings to justify a price, you cannot predict what an agent will do with them.

Badges and scarcity messages mostly only broke ties. They nudged when everything else was equal. They did not drive the decision.

Why free shipping travels well

Free shipping is a number. An agent can see the fee, remove it and compare totals. No gut feeling needed.

Think about a free-shipping threshold. Spend a bit more, save the delivery charge. That is a sum. A machine can do the sum without any nudge.

The research points to a simple rule. What a machine can compute is what a machine can value. The study suggests free-shipping thresholds and similar incentives may carry over well to agent-led shopping.

Note the word "may". This is controlled testing of AI models, not live shoppers with live baskets. Treat it as a strong hint, not a verdict.

The timing matters, because peak season is close. 43% of holiday shoppers plan to start shopping before November, so your offers need to be clear early on.

Deloitte found 38% of surveyed shoppers are 'holiday value seekers'. They shop promotions and redeem rewards. They also expect to spend 12% more than other shoppers.

These shoppers are not only after the cheapest thing. They want a purchase that feels worth the price. A plain free-shipping threshold does that job for people and for agents.

There is a margin side too. Our piece on free shipping, its hidden costs and its real benefits, is a useful companion.

Why psychological pricing loses its grip

A related study from the same author tested pricing tricks on machines. Four AI models used as shopping agents ignored the left digit of 9-endings. A $9.99 price tag got no special treatment.

Partitioned pricing means splitting a price into parts, such as a base price plus a handling fee. The models correctly summed partitioned prices and preferred a single all-inclusive price when the totals were equal.

They also chose the lower price on a one-cent gap every time. Pennies counted. Polish did not.

Position was another story. List position was decisive between comparable options, and one model paid $2 more depending on position.

So charm pricing and drip-style fees lose their pull. A promotion aimed at an agent needs real value in the numbers. Framing will not carry it.

Deeper discounts will not buy visibility

The obvious reaction is to discount harder. The data says that can fail.

In two pre-registered studies, a lesser-known retailer never reached the agent's top 5 under a strict well-known-merchant rule, even at a 60% discount. Those studies used 17,000 responses and three AI models.

Another rule did more damage. A rule against prices outside the normal range filtered out many listings offering 40%+ discounts. Lesser-known retailers lost out more.

The agent's own rules can outweigh your offer. A bigger discount then buys you a smaller margin and little else.

That is a useful reminder for smaller brands. Visibility depends on being known and on prices that look normal. Keep your prices inside a believable range.

This fits something we have said for a while. Blanket discounting is a blunt tool. Shopping agents can smell a discount, as one of our earlier posts put it.

And it was a blunt tool before agents arrived. Les Binet's analysis, reported by Marketing Week, found that about 84% of price promotions are unprofitable.

What to do this week

None of this needs a rebuild. A few checks will do.

• Check your free-shipping threshold. Say it plainly on product and basket pages, with the total cost in view.
• Test your threshold level. Try a few and compare basket size, not just conversion. Small moves show where shoppers add one more item.
• Show one clear price. Fold fees in so the total is easy to read.
• Treat badges and countdowns as tiebreakers. Do not make them your main reason to buy.
• Keep deep discounts for shoppers who need them. A blanket 40% cut does not buy agent visibility.
• Check what worked. Use control groups to measure incrementality, so you know an offer changed behaviour rather than simply handed over margin.

The same logic holds for people. Plenty of shoppers would buy without a discount. Some need a push. Showing targeted offers only to shoppers who need an incentive protects margin and lifts incremental revenue.

That is where Tempo, our predictive intent engine, comes in. It forecasts intent from shopper behaviour. Precovery then acts before a shopper abandons, rather than after.

(The models in this research are generative AI acting as buyers. Tempo is predictive. Different jobs.)

Prediction beats reaction. Fewer, better-targeted offers also mean spending less on promotions overall. It is not the same budget spread around more cleverly.