01 The number of the week
Comscore reports AI Overviews on 39% of desktop searches
Two in five searches on a computer now open with Google’s own answer, before any link to your business appears.
The most quoted figure in AI search this month reached us through a trade recap, with no panel size and no method attached.
Comscore is a measurement company that tracks what people do online through panels of monitored devices. Its latest AI report, covered on September 23, puts Google AI Overviews on 39% of United States desktop searches. An AI Overview is the block of written answer Google prints above the list of links.
The consequence is plain. For about two in five searches made on a computer, the first thing a person reads is Google’s summary rather than anyone’s website. The links did not vanish, but a summary the reader accepts is a summary that ends the search.
Two limits belong beside the number and neither was in the coverage. The figure describes desktop searches only, which is not where all searching happens. No panel size, date range or method was published in the write-up we could read, so the precision of 39% cannot be checked from outside it.
The direction holds anyway, whether you run an IT services firm, a charter operation or a veterinary referral hospital. Even at 30% the answer box is an ordinary part of a search now, and a business named inside it is being described to someone who may never click anything at all.
What to do about it
Search your own brand name and your two most common service questions in a desktop browser, signed out. Note whether an AI Overview appears and whether your business is named or linked inside it. That check takes two minutes and tells you more than any national percentage, because it is your market rather than a panel.
Forever Cited has every commercial reason to repeat a number this large, since a brief about AI visibility reads better when AI visibility is everywhere. That is exactly why it is worth saying that the panel behind it was not published and that we are quoting a trade recap rather than the report itself.
Source
Search Engine Roundtable, Barry Schwartz, “Daily Search Forum Recap: September 23, 2026,” September 23, 2026 · seroundtable.com
02 The evidence
Wenqing Wang’s team tested whether AI can trace its sources
The assistant describing your credentials gets the sourcing wrong about once in four tries and sounds equally certain either way.
Across 123 expert written tasks the strongest of fifteen model setups produced a fully traceable financial answer 71.54% of the time.
A benchmark called FinFIRST was posted on September 21 to arXiv, the open repository where researchers publish before peer review. It asks a narrow question: when an AI search agent answers a financial question, can it show where each part of the answer came from? The team built 123 expert authored tasks from an 18 field taxonomy and a registry of 138 financial sources, with contributions from over 50 finance experts.
Fifteen model configurations were tested. The highest strict pass rate was 71.54%, meaning the best setup produced an answer that was both correct and fully traceable in a little over seven cases in ten. A different model scored highest on individual facts, at 87.59% on what the paper calls the atomic score.
The finding worth carrying is the gap between those two numbers. 17.48% of correct answer attempts lacked a complete, verifiable evidence chain, which is to say the answer was right and the trail behind it was not there. For a wealth advisor, an accounting firm or a tax practice, that is the difference between being described accurately and being described defensibly.
The authors state the limits themselves and they matter. With only 123 tasks after quality control, the paper calls its own dataset relatively modest for broad generalization, results came from one run per model, and the judge saw only final answers rather than the steps behind them. This measures financial search agents, not how an assistant names your practice.
What to do about it
Ask an assistant a question a prospective client would ask about your field, then ask it where each claim came from. If it cannot point to a page of yours for the parts that concern you, the gap is in what it can find rather than in what you know.
For regulated readers the compliance point is separate and sharper. A wealth advisor under the SEC Marketing Rule, an accountant under the AICPA Code or a tax practitioner under Circular 230 answers for claims about their services no matter which machine phrased them.
Source
arXiv, Wenqing Wang and colleagues, “FinFIRST: Benchmarking Search Agents for Financial Information Retrieval, Sourcing and Traceability,” September 21, 2026 · arxiv.org
03 The platform
Google began its fourth spam update of the year
A ranking change that normally finishes in two days is taking two weeks, and Google says recovery can run for months.
The September 2026 spam update is global, covers every language, and started rolling out on Thursday.
Google began rolling out the September 2026 spam update on September 24. This is the fourth spam update of the year, and the previous three took only a couple of days to roll out. This one may take up to two weeks to complete.
A spam update is Google re-applying its own definition of low quality or manipulative content across everything it has indexed. The update is global and is expected to affect all languages, so a change in how your pages are judged can arrive without anything on your site having changed.
The recovery timeline is the part to plan around. Google says it can take many months to recover from a spam update. That is not a penalty you appeal but a reassessment you wait out while doing the work that changes the assessment.
The most exposed businesses are the ones publishing at volume. An IT and cybersecurity firm or a software company with a large library of near identical explainer pages is likelier to be re-sorted than a practice with twenty carefully written ones. If your traffic moves over the next two weeks, note the date before concluding anything about your website.
What to do about it
Write down today’s date and your current search traffic before the rollout finishes. If your numbers move over the next two weeks you will be able to tell a Google reassessment apart from something you changed, which is the most useful thing to know in that situation.
If the move is downward, the answer is to improve the thin pages rather than to publish more of them. Adding volume is what this class of update is built to catch.
Source
Search Engine Roundtable, Barry Schwartz, “Google Releases September 2026 Spam Update,” September 24, 2026 · seroundtable.com
04 The proposal
Four researchers proposed auditing AI sources before answers ship
Someone is finally building a way to check an AI’s sources, and so far it has only been tested against invented records.
The framework reproduced all 81 of its logic combinations and rejected 192 deliberately broken records, none of which came from a real AI answer.
Kainan Zhou, Chuhong Xu, Gangzhen Qian and Zhaoyi Li posted a framework to arXiv on September 24 that scores the risk in a source before a generative answer is built on it. The idea works as a gate: a claim is checked against the standing of its source, and a source that fails is not used.
The validation is where honesty is required. The framework reproduced all 81 three state predicate combinations and rejected 192 deliberately malformed records on an exhaustive synthetic suite. Synthetic means the records were built for the test, and no real AI search answer was evaluated.
That is a proof the logic behaves as specified, which is a real and necessary thing. It is not evidence that AI answers about your business would be judged better under it. A specification that behaves correctly on invented data has cleared its first gate and none of the ones after it.
The reason to report it is direction. Two separate papers this week are about checking the evidence behind an answer rather than improving the answer, which is where attention moves once a system is used enough to be worth auditing. For an attorney or an admissions consultant whose field runs on verifiable claims, a search system that starts grading sources is worth watching.
What to do about it
Nothing today, and that is the honest answer. File it as the second signal this week that the evidence behind an AI answer is becoming something systems try to measure.
The useful preparation is unchanged and unglamorous. Make the claims on your own pages ones a stranger could verify, with dates and named sources, because every proposal in this area rewards exactly that.
Source
arXiv, Kainan Zhou, Chuhong Xu, Gangzhen Qian and Zhaoyi Li, “Claim-Gated Source-Risk Auditing for Generative Search,” September 24, 2026 · arxiv.org
05 The money
Google set a 3.5 star floor on ad ratings
Slip below a 3.5 average and Google removes the stars from your ads by itself, with no notice and nothing to appeal.
A store also needs at least 100 unique reviews inside a rolling 24 month window before Google will calculate a rating at all.
Google published the requirements for store ratings in Merchant Center on September 22. A store must maintain an average composite score of at least 3.5 out of 5 stars, and generally needs at least 100 eligible, unique reviews before a rating is calculated.
The review window is the part most people miss. Google evaluates reviews gathered within a rolling 24 month window, so a review ages out of the calculation. A business with 120 reviews collected three years ago and none since has, for this purpose, no rating.
The consequence is automatic. If the average dips below 3.5 the star rating assets stop displaying alongside your ads until performance improves. Nobody calls, and the ad that was carrying social proof yesterday is carrying none today.
This item is narrower than the others and it is worth saying so. It governs Merchant Center store ratings, so it reaches you only if you advertise with a store attached, which is more of the day for a builder selling products than for a surgical practice. The transferable point is that a trust signal you do not own can be switched off by a threshold you did not set.
What to do about it
If you run ads with star ratings, count how many of your reviews arrived in the last 24 months rather than in total. That is the number Google is using, and a steady flow of new reviews matters more than a large historic pile.
How you may ask for those reviews is not a free choice. Your own regulator sets the limits, and a medical, veterinary or legal practice has tighter ones than a builder does.
Source
Search Engine Roundtable, Barry Schwartz, “Google Merchant Center Requirements For Google Store Ratings,” September 22, 2026 · seroundtable.com
