The Answer Engine Quotes the Sentence, Not the Picture
Google rebuilt its search box in May, the first change in 25 years, and the new surface does not rank pages for you to choose between. It lifts a sentence and speaks it. The data behind that shift is uncomfortable for anyone whose work is visual, and it points at a smaller, more specific job than publishing more.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page

Ask an AI answer engine which shops in Saigon rent a 35mm film scanner and it will name three, add a forum thread from 2024, and warn you that one of them closes on Sundays. Ask which image tool holds a character's face steady across twelve frames and it will quote a paragraph from a blog post you may have read and disagreed with.
That second answer is the one worth sitting with. The engine does not hand you a list of pages and let you judge. It assembles one answer from several sources, keeps the links in, and writes it in its own voice, and the part that gets noticed is the sentence it lifted whole from someone else's page. The same collapse of the click is already visible in referral data, which I wrote about when Reddit lost a million search referrals a week.
The box stopped being a list
According to the Italian newsletter IA per tutti, which interviewed Google Italia sales director Elisabetta Taverna at Intersections 2026 in Milan, Google rebuilt its search bar in May for the first time in 25 years. The field is wider, it accepts images and files, and it suggests how to finish your question instead of waiting for three keywords. AI Mode sits beside it: you ask the way you would ask a person, and it assembles an answer from several sources with the links kept in.
The interview reports two numbers that explain why this changes the job. Queries are now two to three times longer than they were, and one AI Mode query in five starts from a voice clip, a photo or a file. The example in the piece is a person who used to type "men's trail shoes" and now writes: I run three times a week on dirt, I have a wide foot, I want to spend under 150 euros, what do you recommend. The answer cites a specialist shop, a buying guide, a YouTube video and a coach on Instagram, because each one covers a different part of the question.
Taverna's summary of what to do about it is blunt: a good SEO is a good GEO. Same foundations, content that actually helps the person searching. Google's own guidance in the piece is almost boring. Put the important information in text, link internally to the right pages, use structured data so a number is labelled as a price.
The less boring question is what the answer lifts, and which signals decide whether your name is in it at all.
What tracks with being named
The interview cites an Ahrefs study of 75,000 brands, and the study is public, so I read it rather than taking the summary. Ahrefs measured how a set of signals correlated with a brand being mentioned in Google's AI Overviews. Branded web mentions came out at 0.664 on the Spearman scale, branded anchor text at 0.527, branded search volume at 0.392, and Domain Rating at 0.326. Raw backlink count was 0.218. Number of pages on the site was 0.17, which Ahrefs describes as almost no relationship between how much you publish and whether an AI answer names you.
A later Ahrefs study, which I have only seen summarised second-hand, puts YouTube mentions at about 0.737, the strongest signal in the set. Treat that one as a lead rather than a settled figure.
Three caveats before anyone turns this into a strategy deck. Correlation is not causation, and Ahrefs says so in the post. The sample only includes domains with a Domain Rating above 40, which excludes most small shops and every new store. And the study counts mentions of a brand, not citations of a URL, so it tells you who gets into the room, not which page gets quoted once the answer is being written.
Even with those limits, the shape is hard to argue with. Language models are text machines. They know a brand because the brand's name appears in sentences, more than because a page about it has links pointing at it. That is why the weakest number in the set belongs to the thing content teams have spent a decade doing: publishing more.
It is also a slightly awkward result for a blog. If page count tracks so weakly with being named, then publishing more posts is not an obvious lever, and being named in other people's sentences looks like a stronger one. That is a hypothesis drawn from an association, not something the study tested as an intervention, and it is the thing I would want to test on my own pages. Better to write it down that way than pretend the data says something friendlier.
A page has to be readable before it can be quoted
Adobe's AI traffic report from August 2026 has the number I keep coming back to. Its Citation Readability Score measures how much of a page's content is effectively visible to AI, and the homepage scores in that report run from 44 percent for general merchandise to 74 percent for cosmetics, with furniture and apparel at 61. At the bottom of that range, more than half of a homepage never reaches the model, even though a person with a browser sees all of it.
Think about what sits outside those scores. The price, in an image. The sizes, in a widget that renders after the fact. The delivery window, in a banner. The licence terms, behind a tab. A model reading the page gets the text it can parse, and the price may not be in that text.
Adobe's own read of the pattern is consistent across industries: blogs, guides, FAQs and support pages score highest, while booking, checkout and account-creation flows score lowest regardless of vertical. Structured, informational pages get read. Flows assembled from interactive widgets are harder for a model to read, which is a weaker claim than unreadable and still the difference between a price that is in the text and a price that renders into place a second later.
The same report has the numbers that keep this from being a thought experiment. AI-referred traffic to US retail sites was up 62 percent year over year in July 2026. Those visitors converted at a rate 60 percent higher than non-AI traffic, generated 53 percent more revenue per visit, spent 59 percent more time on site and bounced 33 percent less. Adobe also found that 89 percent of consumers say they click the links or sources an AI assistant gives them, and 66 percent go to the brand's own site to verify the recommendation.
Two honest caveats. This is vendor data from one analytics company, covering US retail, not creative tools. And AI referrals are still a small slice of overall traffic; Adobe's vice president of strategy described LLM-based purchasing as early-stage when the report came out. The conversion premium is real and the volume is not yet the main event. Plan accordingly.
The test takes ten minutes
The newsletter's own exercise is the most useful thing in it, so here it is in English.
Open AI Mode in a private window, with your preferences and history out of the picture. Ask three questions.
- The generic one. What a stranger would type: "ai image tool for product photos."
- The specific one. What your best customer would type, with the constraints spelled out: "I sell handmade ceramics on Shopee, I photograph them on a phone against a white wall, I need consistent lighting across forty listings and I do not want a subscription."
- The comparison. "X versus Y for someone who mostly edits existing photos."
Write down who gets cited and which sentence gets lifted. Do it again in a month. The newsletter's author reports that changing one factor in the test was enough to move their own site from absent to cited, and does not present that as a controlled experiment. Treat it as one anecdote about query phrasing. The Ahrefs numbers above are about brand signals, not about whether a detailed question retrieves a specific page, so that question is still open. What I would test is whether a page written for one narrow situation gets lifted where a general one does not. It is the same move as decomposing a prompt into decisions instead of stacking adjectives, except the reader is a retrieval system rather than an image model.
If you are absent from all three, the problem is probably not your SEO plugin.
What I would change first if I sold images
Pick a territory narrow enough to be the answer to something. The interview mentions Enel, which stopped defending individual keywords and started working on customer value over time by pooling data across marketing, the store and the call centre. You can run the small version of that. Decide which specific situation you want to own, and write for that situation only.
Then make the page legible. Price, format, file size, turnaround, licence, what it will not do. In text, not only in a screenshot. A model can pull text out of an image, but it does so less reliably than it reads marked-up prose, and it cannot quote a layout. It can quote the sentence that says what the image is for, who it suits and what it costs.
Then get the name into other people's text. Transcripts, roundups, forum answers, a video where you say the tool names out loud, since YouTube mentions came out strongest in Ahrefs' follow-up. This is slower and less controllable than publishing, which is presumably why the content-volume strategy survived so long.
And if you sell on a marketplace, treat the listing title and description as the retrieval layer, not decoration. Write them in the buyer's language, with the material, the size and the use case spelled out. The photos close the sale. The words are what get the listing into the conversation in the first place.
The part that stays the same
A model repeating your sentence does not mean the person arriving will stay. The craft still decides that, and the numbers above are about getting into the room, not about what happens once someone is standing in it.
But the room has changed. The picture still sells the work. The sentence is what gets the work quoted, and most of us have spent years treating the sentence as an afterthought.