01 The new measurement
Google starts reporting the searches made with a camera
You can now count the customers who found you without typing a word, though Google will never tell you what they wanted.
Search Console began separating image searches from typed ones globally today, and the query column stays blank for every one of them.
Google Search Console is the free dashboard telling a website owner which of their pages showed up in Google and how often. As of this morning it separates searches made with an image from searches made with words. Google’s own announcement says the integration of multimodal search data is rolling out globally starting today.
These are searches your customers already make without thinking of them as searches. Google Lens, Circle to Search on Android, an image uploaded to Google Search, and the Chrome right-click “Search this image” all count. Someone photographs a roof detail, a part number or a building, and your page is what comes back.
There is one thing you will not get, and Google states it plainly in the documentation. Because these searches mostly use images rather than text, “specific text query data isn’t available for this traffic” and “the queries dimension isn’t available when this search type is selected”. You can see which of your pages appeared and how often, and never what the person was actually looking for.
John Mueller of Google summarized it the same morning: you do not see the photo, but you do see which of your pages showed up and how often. Barry Schwartz noted the gap that remains, which is that there is still no click or query data in the report covering generative AI features.
What to do about it
Open Search Console, choose the Search type filter on the performance report, and look for the new Text-based and Multimodal options under Web. If multimodal impressions are arriving, the pages they land on are ones where your images are doing work you were not crediting them with. An architecture studio, a historic restoration specialist or an orthopedic practice will see this before an admissions consultant does, because their customers point cameras at buildings, materials and injuries.
Source
Search Engine Roundtable, Barry Schwartz, “Multimodal Search Type Filter In Google Search Console Performance Report,” September 24, 2026 · seroundtable.com
02 The evidence
Zhaohui Wang tested AI recommendation without supplying the answer
Benchmarks showing AI recommends well usually hand it the right answer first, and taking that away cuts the score roughly in half.
Across three Amazon datasets the usual testing protocol overstated performance by 92 to 95 percent, and eight separate attempts to improve on a plain baseline all failed.
Most tests of whether an AI recommends well are run under what this paper calls an oracle protocol, meaning the correct item is always present in the list the model is asked to sort. Zhaohui Wang posted a study to arXiv on August 27 that took it out. Across three Amazon datasets, the usual protocol overestimated realistic performance by 92 to 95 percent.
The cause sits in the step before the model runs at all. Realistic retrieval covers only 2 to 19 percent of relevant items across eight datasets in three domains, which puts a hard ceiling on anything the ranking model can do afterwards. If the right option was never gathered, no amount of sorting will surface it.
Then comes the part worth sitting with. Prompt engineering, scaling the model across a 168-fold parameter range, sequential models, supervised rerankers, fine-tuning, hybrid retrieval and score-aware prompting were all tested, and none significantly beat a plain collaborative-filtering baseline. Feeding the baseline’s own scores to the model mainly made it reproduce that baseline’s order.
The limits are real and the author states them. This is Amazon shopping data, not the question of which accountant or surgeon an assistant names, it is a preprint, and he writes that the ordering he found need not hold in systems with higher recall. What survives the caveats is the direction: where recall is low, improving what gets gathered matters more than improving what does the ranking.
What to do about it
Take it as a question rather than an instruction, because it measures shopping baskets. The question is whether a system that would recommend a business like yours has anything of yours to gather: a page for each service you offer, a presence on the places your profession is listed, and a name that resolves to one organization rather than five.
Forever Cited makes its case on the claim that being gathered is what decides whether a business gets named, so a study concluding exactly that is one we have every commercial reason to repeat. That is the reason to say plainly that it measured Amazon product data and that its own author flags the limit.
Source
arXiv, Zhaohui Wang, “The Recall Ceiling of LLM Recommendation Reranking,” August 27, 2026 · arxiv.org
03 The structure
Yunfei Zhong’s team says AI search retrieves for answers, not readers
Your page is no longer competing to be read by a person. It is competing to be used by a machine writing the answer.
A framework posted this week proposes judging a retrieved page on its source, its date and whether its facts hold, before any answer is written.
Yunfei Zhong and nine co-authors posted a framework to arXiv on September 20 and revised it on September 23. In ordinary web search a person reads the ranked results and works out the answer themselves. In AI search those same pages are inputs to a model that writes the answer instead.
That changes what retrieval is optimizing for, and the paper names three stages. Answer Support identifies which pages actually contribute information to the answer, and Content Trustworthiness then assesses whether that information is a reliable basis for a correct answer “from source, temporal, and factual perspectives”. Context Organization selects and structures whatever survives, “under a finite context budget”.
Read those three criteria as an owner rather than an engineer: who published this, how recent is it, and does the claim hold. Then, because the space is finite, a page can be accurate, relevant and still left out because something else said the same thing more usably.
The caveat is the size of it. The paper reports “consistent improvements” at both the retrieval and the answer level, and publishes no sample size, no dataset size and no numeric result in its abstract. So it describes how a serious team thinks the plumbing should work rather than measuring how well it does. Read it as a statement of direction.
What to do about it
Put the three questions to your single most important page today. Is the author named and credentialed on the page itself, is there a visible date, and is every specific claim checkable against something outside your own website. An admissions consultant writing about this year’s application timeline and a wealth advisor writing about a contribution limit both fail the date test the moment the year turns, and neither failure announces itself.
Source
arXiv, Yunfei Zhong and co-authors, “From Ranked Documents to Reliable Contexts: An Answer-Oriented Context Construct Framework for AI Search,” September 20, 2026, revised September 23, 2026 · arxiv.org
04 The listing
Google tests letting an AI set your opening hours
You will describe your hours in a sentence, and software will decide what that sentence meant before customers read it.
Google Business Profiles is testing a feature that reads a free-text description of your opening hours and converts it into the structured hours customers see.
Google Business Profiles is the panel carrying your name, address, phone number and hours when somebody searches for you. Google is testing a feature that asks you to describe your opening hours in ordinary writing, then uses AI to turn that description into the structured hours on your listing. The prompt reads “Set your hours automatically. Describe your opening hours. Google will extract your hours and update them below.”
Ayman Ali posted the screenshot on X, and the appeal is obvious to anyone who has clicked through seven days of dropdown menus twice a year. Barry Schwartz reads it as one of the small time savings small businesses appreciate, which is fair.
The exposure is in the word extract. Opening hours are full of things a sentence handles badly and a dropdown handles exactly: closed between one and two, last appointment an hour before the door shuts, Saturdays by arrangement. An interpretation that is almost right publishes as hours that are simply wrong, and the person who finds out is standing outside your door.
This is a test rather than a rollout, spotted on a single screenshot rather than announced, so there is nothing to turn on or off today. The thing worth knowing is that if it does appear in your profile, the sentence you type is a draft and not an instruction.
What to do about it
If the option shows up, write the sentence, then read back what it produced day by day before you save, the same way you would check a form somebody else filled in on your behalf. It matters most where the clock is part of the service. A tax practice whose hours triple between January and April, or an architecture studio open by appointment only, has opening hours that a plain sentence describes badly and a customer relies on completely.
Source
Search Engine Roundtable, Barry Schwartz, “Google Business Profiles Tests Set Your Hours Automatically,” September 24, 2026 · seroundtable.com
05 The money
Google Ads will propose moving budget between your campaigns
Google will offer to take money off your weaker campaigns to feed your stronger ones, and one of the two modes does precisely that.
A Google Ads recommendation now arrives in two named modes, and only one of them adds money without taking any away.
Google Ads has updated the Recommended Investment Strategy on its Recommendations page with two named modes, and the difference between them matters. Holistic mode reallocates existing spend out of campaigns Google judges underused or less efficient and into your top performers. Growth mode adds budget to the top performers and, in Google’s wording, does “not reduce spend on any existing campaigns”.
Google’s help documentation describes Holistic mode as a way to “reallocate existing spend from underutilized or less efficient campaigns”. Hana Kobzóvá, who spotted the change first, reads Holistic as reallocating existing budgets while adding weekly spend, and Growth as strictly adding to campaigns the budget is holding back.
The word doing the work is efficient, because efficiency here is counted in conversions. A campaign that produces few form fills but keeps your name in front of the one referral source that matters is an underperformer by that measure, and Holistic mode is the setting that moves money away from it.
Google is explicit that the choice stays with you. “You retain full control over your account. The tool doesn’t modify excluded campaigns,” and any campaign you untick keeps its budget and bid strategy. Eligibility is limited to accounts where a budget is already constraining at least one performance campaign, and the projections are typically seven-day forecasts.
What to do about it
If the Recommended Investment Strategy appears in your account, read which mode is proposed before you read the number next to it. Growth mode is a spending decision, and Holistic mode is a reallocation decision that reaches campaigns you never asked it to touch.
Untick anything whose job is not counted in conversions, such as a campaign defending your own name or one aimed at a referral audience, before you apply any of it.
Source
Search Engine Roundtable, Barry Schwartz, “Google Ads Investment Strategy New Holistic & Growth Modes,” September 24, 2026 · seroundtable.com
