With the proposal feedback in hand, the plan was simple: ask people, then read until the question held. Both changed the project more than I expected.
Twenty people
I ran a needs-finding survey with 20 respondents, shared through my own network: 12 aged 18 to 24 and 8 aged 25 to 34, and 15 of them had shopped in more than one country. It is a small, formative sample. Its job was to keep the design close to what shoppers say they need.
What they said:
- Help with product information was welcome: 18 of 20 rated it useful or very useful.
- They already reach for it: 13 of 20 use their phone while shopping to look up product information, compare prices or read reviews.
- They want a say: 19 of 20 rated being able to "control, customize, or turn off the information shown" as important or very important.
- The most common problem was finding a specific product, named by 12.
- Asked which kind of store they would want this support in, 10 chose grocery, 5 furniture, 3 both and 2 neither.
One open answer stayed with me. Asked what would make them refuse this kind of technology, a respondent wrote: "If it suggests anything without my asking. Like all the ads we face on screens."
The survey also asked which device people would prefer for this help: 10 said a smartphone, 6 said either would be fine, 3 said smart glasses and 1 said neither. People who have never worn display glasses in a shop cannot really judge them, so I do not lean on that answer in either direction. It does mean the thesis cannot claim that shoppers are asking for glasses.
Two cuts followed. Afterwards I had written that "covering both in one study does not produce twice the knowledge. It produces diluted findings", and the survey leaned the same way, so grocery became the only setting. IoT went too, since the data would have been prepared in advance anyway.
Reading to disprove
Then came the reading, and at some point I changed how I read. I put it like this at the time: "Searching for support produces a reference list. Searching for disproof produces a literature review." Instead of collecting papers that agreed with the idea, I looked for the ones that could break it.
They broke a lot.
The design still compared a whole assistant, with navigation, product information and comparison, against shopping without it. An audit of that design named the problem in one line: "Three-feature bundling limits mechanism attribution." If people did better with the assistant, nobody could say which part did it, or whether it was simply the novelty of wearing a headset.
So the bundle came apart. Wayfinding went first; in my notes I wrote "this is not my research anymore". The health framing went next. I had imagined showing nutrition and allergen information, until I wrote down the obvious: "users can get nutrition from the packaging, so we do NOT show nutritious information." The problem was not missing information. It was getting the right information off a package, in a shop, at the moment you need it.
One note from that period turned out to hold the whole thesis: "UX issue: timing is unresolved."
When should it speak?
Proactive assistants on head-worn displays were already an active research topic, and two systems shaped my thinking. AiGet shows background knowledge on smart glasses while people walk around campus, shop or visit a museum; both AI-started and user-started help are available, but the two were not compared head to head, and the content differed between them. Sensible Agent offers help on a headset and waits for the user to confirm; it was compared with a voice assistant in which users phrased their own questions, so the starting move changed together with the input.
That is where the question formed. Not what the assistant says, but who starts it.
Three things I got wrong
The reading also corrected me, three times. Each correction changed how I work.
A sentence one paper undid. I had written that no study compared a head-worn shopping assistant with a smartphone. Then I read Álvarez Márquez and Ziegler (2023). They compared a head-worn in-store product system with a smartphone app built for the study, with the same functions and information, and found the headset on par in practicality, with possible benefits for discovering products. Their study also shows how tangled such comparisons are: on the phone, the user had to choose a product before recommendations appeared; on the headset, recommendations appeared directly for the product in view. The device and who starts the help changed together.
After that I made a rule for my writing: never claim that nobody has done something. Describe what the nearest studies did and let the open question follow from that. A sentence about three papers I have read can be checked. A sentence about all of publishing cannot, and one paper is enough to undo it.
A survey number that meant something else. An early draft said that 9 of the 20 respondents photograph a package and ask an AI on their phone. When I re-read the survey export row by row, there was no such option. What the data supports is that 13 of 20 use their phone in the shop for product information, prices or reviews. Since then every survey number in the thesis comes from one verified tally of the raw export.
A paper read backwards. I had cited Hoffmann and colleagues (2022) as showing that information pushed without user control hurts purchase intention. The full paper says the reverse. In their field study at the point of sale, when the information was detailed, giving shoppers control over which details they saw made it less persuasive than showing everything at once. Since then I describe a paper's findings only after reading the full text, and each claim in the thesis is tied to a quote from it.
None of these was comfortable. All three made the thesis stronger, because each one replaced a claim I wanted to be true with one I can show.
Where it landed
The question was no longer whether an AR assistant beats shopping without one, or whether glasses beat phones. Both comparisons change too many things at once. It became one design decision inside the assistant: when it has something useful to show about the product in your hand, should it wait to be asked, offer first, or just show it?
The next post is about the study built around that question.