Lightspeed's AI shopping study and what smaller brands should do

AI shopping assistants recommend national chains over smaller retailers by a wide margin, according to a study Lightspeed Commerce published on 29 September 2026. In head-to-head comparisons, ChatGPT and Google’s AI tools picked the larger retailer up to 94% of the time, so specialist and challenger brands need to check how often assistants name them.

Inside the Vaer AI numbers

AI search consultancy Vaer AI ran the research for Lightspeed, which sells point of sale, payments and ecommerce software to retailers. According to Lightspeed’s release, the team ran 20,000 shopping prompts across 10 product categories, set in Los Angeles, San Francisco, New York City and Montreal, through ChatGPT, Google AI Mode and Google AI Overviews between June and August 2026. That produced roughly 460,000 responses and more than 2.4 million links. No prompt mentioned store size, so any lean toward big names came from the assistants themselves.

Answering from training data alone, the assistants chose the national chain over a smaller rival 90 to 94% of the time. Live web search narrowed the gap without closing it, since chains still took 46 to 58% of recommendations with search switched on. Local retailers were the top pick in roughly a third of broad category searches, falling to as little as one in ten once the shopper named a specific product. Google AI Overviews leaned hardest, recommending no local store in 68% of shopping answers, while ChatGPT and AI Mode named a local retailer in roughly 70% of theirs.

Vaer AI then re-ran matched prompts with “independent” added, and the share of recommendations going to large chains dropped from about 44% to as little as 9%. A companion Censuswide survey of 2,000 North American consumers found 56% had used AI to help with a shopping decision, and 50% said they would shop locally more often if AI made independent retailers easier to find.

Why a specialist brand loses the product-level query

Smaller retailers lost most ground on the queries closest to a sale, since a shopper asking where to buy a particular model of trail shoe has usually decided what they want. A chain with national stock and years of reviews and coverage is likely to look like the safe answer to a model trained on the open web, and Vaer AI’s founder Tom Wells said in the release that the bias “begins in training”.

Live search helped smaller retailers because the assistant reads current pages and listings instead of relying on training data. Content and SEO teams can only influence that half, so they should concentrate on the pages assistants fetch, such as product pages, stockist pages and the editorial round-ups beside them.

Because assistants changed their answers on a single word in the prompt, shoppers’ own descriptions of what they want shape which retailers get considered. That makes it a category entry point problem. If your customers choose you because you are independent, specialist, local or expert-led, that attribute has to appear in plain text wherever an assistant reads about you.

Our view at Manual Focus is that challenger retailers and DTC brands should stop treating AI visibility as a single score and track it at product level, because the study shows a brand can look healthy on broad category prompts and still disappear at the point of purchase. Retailers already watching AI referral traffic for peak trading, as we set out in our read of Adobe’s holiday forecast, can add this as the visibility measure that sits upstream of those clicks.

How can a smaller brand get named by ChatGPT?

1. Write 30 product-level prompts. Use your ten best sellers and phrase three buying questions for each the way a customer would, including your town or region where location matters.
2. Run them in pairs. Ask each prompt plain, then again with the attribute that sets you apart, such as “independent” or “specialist”, in ChatGPT, AI Mode and AI Overviews. Record who gets named in each version.
3. Put the attribute where assistants read. Add it to the homepage, about page, product page copy, Google Business Profile and Merchant Center feed in plain words rather than a logo or image.
4. Find the pages assistants cite. Note which round-ups, review sites and directories appear in the links for the prompts you lose, then brief PR on getting listed in them.
5. Repeat monthly into peak. Re-run the same set in early November to see what moved before Black Friday.

Why the study needs a second opinion

Lightspeed commissioned the study, sells software to the smaller retailers it describes and used the release to launch an AI visibility campaign aimed at them, so it has a commercial interest in the finding. The release gives headline figures and a methodology note but no full dataset or category breakdown, and an AI classifier sorted each retailer as large, mid or small. The prompts were set in four North American cities and the survey covers North American shoppers only, so there is no UK figure yet. The results also reflect the versions of each assistant that were live between June and August. None of it explains why a given brand was left out, which only your own prompt tests will show.

If you want help setting up a product-level AI visibility audit, talk to us.

I run marketing for [brand], [one line on what you sell], at [website]. We
sell mainly in [market], and customers choose us because we are [independent /
specialist / local / expert-led - pick what is true].

Attached (or pasted below) is our list of our ten best-selling products, with
any real customer questions we have from reviews, search queries or the help
desk: [attach or paste].

For each product, write three buying questions a shopper would type into an AI
assistant when they have already decided roughly what they want, phrased the
way a customer would say them, not the way we describe the product. Where
location matters for buying, include [town or region].

Then write a paired version of every prompt that adds our distinguishing
attribute ("independent", "specialist" or whichever applies), changing nothing
else.

Give me a table with the columns: Product | Prompt number (1-3) | Plain
prompt | Paired prompt (with attribute) | Why a shopper would ask this.

That should be 30 rows. Don't invent product details that aren't in what I've
given you; if you need one, put it in [brackets] for me to fill in.

Set this up as a routine that runs on the first Monday of each month, with an
extra run in early November before Black Friday.

Here are our 30 test prompts, each in a plain and a paired version: [paste the
table from the first prompt]. Our brand is [brand] at [website].

For each prompt and each version, check the answer in ChatGPT, Google AI Mode
and Google AI Overviews [from a location in our market]. If you can't reach
one of those assistants yourself, tell me which, and I will paste its answers
in for you to log.

Record which retailers or brands each answer names and links to, in order.

Give me a table with the columns: Date | Product | Prompt | Version (plain or
paired) | Assistant | Retailers named, in order | Are we named? (yes/no, and
position) | Large national chain named first? (yes/no).

Then give a short summary: how often we were named overall, plain versus
paired, broad versus product-level, and what changed since the last run.
Where an assistant showed no answer or wouldn't load, say so in that row
rather than leaving it blank.

Here are the prompts from our latest test run where [brand] was not named,
with the links each AI answer cited: [paste the rows and links].

Go through every cited link and sort them into: editorial round-ups and buying
guides, review sites, directories and listings, retailer or brand pages, and
other.

Give me a table with the columns: Site | Page title | URL | Type | Number of
lost prompts it appeared in | Which products it covers | Are we mentioned on
the page? (yes/no/couldn't check).

Rank by the number of lost prompts. Then list the top ten editorial, review
and directory pages as a brief for our PR team, with one line on what each
page seems to need to include us. Mark any page you couldn't open, and don't
guess at its contents.