Bain's holiday forecast came out this month with a number that is easy to skim past. Twenty-four percent of US online shoppers say they will start their holiday shopping on an AI platform this year, up from seventeen last year. That is not most people, and most purchases will still finish on a retailer's site. But the first step of a quarter of journeys now happens somewhere that behaves nothing like a search results page, and where very few brands know whether they appear at all. What follows is what the number means, what it does not mean, and what can realistically be done before Black Friday.

Key takeaways

Bain expects 24 percent of US online shoppers to start the season on an AI platform, against 17 percent last year. Sixty percent will still start on a retailer's own site.
An AI answer returns three or four product names rather than a page of ranked links. A brand that is not named is not under consideration.
Only 16 percent of brands measure whether they appear in AI answers, and the tools that have grown up around the problem do not agree with each other.

What Bain asked, and what the number does not say

Bain's Consumer Lab surveyed more than 1,100 US consumers for its annual holiday forecast. Twenty-four percent said they expect to begin their holiday shopping on an AI platform such as ChatGPT, Google Gemini or Claude. Sixty percent said they would begin on a retail or brand website, up from 51 percent last year, and a further 13 percent said they planned to use a retailer's own AI assistant.

Both of the big numbers went up at once, which tells you people now start in more than one place. It also means the 24 percent is not a share of purchases. It is a share of people describing where their journey begins, and stated intent usually runs ahead of behaviour, especially for anything that feels new.

So the honest reading is narrow. A quarter of shoppers expect to open the season by asking a model, and then carry on somewhere else. That first conversation is short, it happens before any comparison, and it is the part of the journey brands currently see the least of.

An AI answer is a shortlist, not a ranking

This is the part that matters more than the percentage. A search results page shows ten links. You can rank fifth and still get traffic, and the shopper does the comparing. An AI answer names three or four products and explains why. There is no second page, no position eleven, and nothing to scroll past.

The consequence is blunt. If the model does not name you, you are not in the consideration set. Not ranked low. Absent.

What decides that is not your homepage. Models answer from what they have absorbed and from sources they can retrieve and parse quickly, which in practice means your product data, retailer listings, review sites, spec sheets and press coverage. Your own marketing site is one input among several, and often not the one quoted.

That also explains something that has been confusing people all year. Visibility in conventional search does not predict visibility in AI answers. Analyses of Google's AI Mode have found that for products listed by several sellers, the seller shown first differs from conventional search results close to half the time.

Almost nobody is measuring this, including the people selling the measurement

The Interactive Advertising Bureau, the trade body behind most of the measurement conventions the advertising industry runs on, published the first shared vocabulary for tracking brand visibility in AI answers in August. It is called Measuring Visibility in the AI Era, it runs to 36 pages, and it found that only 16 percent of brands systematically track any of this.

It also found more than twenty companies selling tools to measure it, using different question sets and different scoring, which is why two vendors can report different share-of-voice figures for the same brand in the same week. The IAB deliberately avoided calling the document a standard, on the grounds that you cannot standardise a moving target. The framework covers organic mentions only. Paid placements are listed as future work.

Meanwhile the category is being priced as though the problem were solved. Profound, whose product does nothing but report whether a brand appears in AI answers, raised at a 1.8 billion dollar valuation this month. Deloitte told clients this week that generative engine optimisation, the practice of getting named by AI models the way search engine optimisation gets you ranked, has arrived and that most brands have not noticed.

There is a cheaper first step than buying a dashboard. Write down the thirty questions a customer actually asks before buying in your category. Ask each of them three times, on two or three different assistants, and record what comes back. It takes an afternoon, it costs nothing, and it gives you a baseline you understand. Model answers vary between runs, which is precisely why one query asked once tells you nothing.

Most of the work does not fit in nine weeks. Three parts of it do.

Building the authority that makes a model quote you is a matter of quarters. Restructuring a site, earning coverage and fixing a catalogue are not nine-week projects, and anyone selling them as such is selling something else.

Three things can be done in the time left.

Fix the product data on the products that matter. Not the whole catalogue. The twenty or thirty items that carry the season. Complete specifications, real dimensions, materials, compatibility, what is in the box. Models reward information they can state without hedging, and drop products they cannot describe precisely. If that information is scattered across a shared drive and three spreadsheets, the problem sits upstream of AI, and the DAM workflows that actually scale deal with it separately.

Correct what other people say about you. Retailer listings, marketplace pages, review sites and spec aggregators are what the model quotes. If a retailer has the wrong dimensions or an outdated price, that is the version the model repeats.

Answer the buying questions somewhere readable. The comparisons, the sizing, the compatibility, the returns. Put them in plain text on your own pages, phrased the way customers ask them rather than the way the brand describes itself.

The paid route is open, and newer than it looks

Sponsored placements now appear inside AI answers. Comscore, which began measuring them this month, found that on hotel queries in ChatGPT the share of responses carrying a sponsored placement went from 6 percent in March to 24 percent in May. Amazon opened its advertising platform to ChatGPT inventory in September, so brands already buying through Amazon can extend campaigns into ChatGPT as a managed service without opening an account with OpenAI.

Two things are worth knowing before the budget moves. It is not search advertising. There is no keyword auction you control, and OpenAI decides where and when an ad appears. And independent measurement of the paid layer is a few weeks old, which means you will be reporting on it with tools that did not exist at the start of the summer.

Buy it as a test with a stop date and a threshold. Not as a channel.

The case that this is overblown

Sixty percent of shoppers still say they will start on a retailer's website, and that number rose faster this year than the AI figure did. Amazon remains where most people go first. A journey that begins with a question to ChatGPT very often ends in a marketplace search two minutes later, and nothing in Bain's data tells you how much revenue actually turns on the first answer.

It is also worth being clear about who is raising the alarm. The tooling category just raised at more than a billion dollars, consultancies are launching GEO practices, and agencies report demand for this work running ahead of demand for classic search. That is an incentive rather than a conspiracy, but it should change how you read the urgency.

The reasonable position sits in the middle. This matters most where purchases are considered and specification-heavy, in electronics, appliances, travel, tools, B2B equipment. It matters least where purchase is habitual or impulsive. If you sell chewing gum, the shortlist is not your problem this year.

What to do, depending on your role

First, for marketing and brand directors. Run the thirty-question test yourself before commissioning anything. You need to know whether you have a visibility problem before you buy a visibility product, and the answer is often category-specific rather than brand-specific. If nobody in your category is being named consistently, the opportunity is larger and cheaper than any vendor will tell you.

Second, for ecommerce and performance leads. Treat sponsored placements in AI answers as an experiment with a fixed budget and a fixed end date this season. The inventory is real and the audience is real. The measurement is not settled, and committing an always-on budget to a channel you cannot yet report on properly is how a good test turns into an awkward January.

Third, for whoever owns product data. In most organisations that is nobody in particular, which is the actual finding here. Specifications live in a spreadsheet updated at launch and never since, while the key visual goes through four rounds of approval. That imbalance was survivable when a human read the page. It is not survivable when a model decides whether to name you based on whether it can describe you. The same gap shows up every time a team rebuilds an asset that already existed because nobody could find the original.

That last point is where this lands on creative operations rather than on media. The asset and the data describing it now do the same job, and in most production chains they travel separately. The file goes through review while the specification sits in a spreadsheet with no owner. MTM keeps both in the same place, with assets organised automatically and findable in plain language rather than through tags somebody has to remember to apply.

See how MTM handles assets and the data attached to them → https://www.mtm.video/platform/

FAQ

What does Bain's 24 percent actually measure? It measures stated intent, not purchases. Bain's Consumer Lab asked more than 1,100 US consumers where they expect to begin their holiday shopping, and 24 percent named an AI platform such as ChatGPT, Gemini or Claude, up from 17 percent a year earlier. Sixty percent named a retail or brand website. The figures overlap because people start in more than one place, and none of it describes where the money eventually lands.

Is generative engine optimisation just SEO with a new name? It shares the goal and very little of the method. Search optimisation aims at a ranked position on a results page, and rewards keywords, links and page structure. Generative engine optimisation aims at being named inside an answer, and rewards complete product data, sources a model can parse, and third-party coverage it trusts enough to quote. The practical difference is that half the work sits outside your own website, in retailer listings and review sites you do not control.

Can we buy our way into AI answers? Partly. Sponsored placements now appear alongside AI responses, and Amazon opened its advertising platform to ChatGPT inventory in September so existing Amazon advertisers can extend campaigns there as a managed service. The placement is an ad unit next to the answer, not a change to the answer itself, and the platform decides when and where it runs. Independent measurement of these placements began only this month.

How do we find out whether we are visible without buying a tool? Write down the thirty questions customers ask before buying in your category. Ask each one three times across two or three assistants, and record whether your brand is named, where in the answer, and how it is described. Repetition matters because the same question produces different answers on different runs. This gives you a baseline in an afternoon, and it is more useful for deciding what to do next than a single vendor score.

We sell through retailers rather than direct. Does this affect us? More than it affects direct sellers, because the sources a model quotes about your products are mostly your retailers' pages. If a retailer lists the wrong dimensions, an outdated price or an incomplete specification, that is what gets repeated in the answer, and you will not see it happening. The practical step is to audit how your top products are described on the three or four retailer sites that carry the most of your volume.