01 The market data
Google now answers the questions under your results itself
The People Also Ask box used to send readers to whoever answered best. In the first week of September, 97% of those answers were Google’s own.
Below almost every Google result sits a box of follow-up questions, the People Also Ask panel, and for years it was a small doorway: click a question and a snippet from somebody’s page opened, with a link to it. Search Engine Roundtable reported on Wednesday that the doorway has closed. In the first week of September, 97% of People Also Ask answers were AI Overviews, written by Google rather than lifted from a page, according to AlsoAsked’s analysis of 19.2 million English-language queries.
The speed of the change is the story. The share was 86% in August and about 12% fourteen months ago. A second tracker, Allintitle’s Also Ask Miner, reports the figure reached 100% from August in its own data.
Mark Williams-Cook of AlsoAsked says the numbers put it “on track for 100% shortly.” Both companies sell People Also Ask tracking, which is worth knowing; the two datasets were built separately and land in the same place.
A day later Google gave the change its official framing. Asked on Reddit about tracking rankings, John Mueller of Google wrote that “the old ‘position 1 - 10’ is hard to map, or to make useful for site owners.” Put the two together and the picture is plain: the questions a prospective client asks after the first result are now answered by Google, and Google itself says the old rank report no longer describes the page.
For a professional firm, those follow-up questions are where the specific work lived. “How long do I have to file,” “what does a consultation cost” and “can I do this myself” were the queries a well-written FAQ page could own outright. That page still feeds the answer. It just stops receiving the visit.
What to do about it
Pull your Search Console data for the queries that are questions (they start with how, what, can or does) and compare clicks this September with last. If impressions held and clicks fell, you are being read and not visited, which is the pattern above and not a ranking loss.
Keep the FAQ pages. Rewrite each answer so the first sentence is the complete answer with your firm named in it, because that is the sentence an AI Overview lifts, and stop reporting average position to anyone as evidence of anything.
Source
Search Engine Roundtable, “Almost 100% Of People Also Ask Results Are Now Answered By Google AI Overviews,” by Barry Schwartz, September 9, 2026 · seroundtable.com · Google’s comment: Search Engine Roundtable, September 10, 2026
02 The evidence
The AI got the law right and the citations wrong
A new paper tests six AI systems on 185 legal questions and finds a correct conclusion can sit on citations that do not support it.
A paper posted to arXiv on Tuesday opens with a sentence every lawyer will recognize: “a language-model answer is only useful to the extent that a reader can verify each claim against the source the system cites.” The authors, a team led by Chen Qian, argue that the usual way of grading AI answers scores the answer as a whole, so “a correct conclusion can rest on fabricated or loosely matched citations and still score well.”
So they graded the citations. On a 185-question legal benchmark drawn from LegalBench, spanning constitutional law through tort, they ran six systems on the same underlying model with the same retrieved documents and the same instruction to cite. An answer counted only if a judge scored it correct and every citation resolved to a passage the system had actually retrieved, which they call strict accuracy.
Under that standard the numbers fall hard. Plain prompting managed 38.4%. A standard retrieval pipeline with re-ranking reached 43.2%, and two popular agent frameworks scored 21.6% each. The best baseline, a self-checking method, reached 59.5%.
The authors’ own system, GANDR, which adds a second agent whose only job is to audit each claim against its cited source, reached 70.8%, and against two law-trained annotators its audit caught under-supported claims at an F1 score of 0.84.
Two cautions: this is a preprint, not peer-reviewed, and the authors are grading their own design, so the 70.8% is the best case for their method rather than for AI legal answers generally. The finding that survives both is the gap itself: on a legal question, the right conclusion and a citation that supports it come apart roughly a quarter of the time even in the best system tested, and far more often in the tools most people use. It is the brief that reaches the right holding on pin cites that do not say what it claims.
What to do about it
When a client arrives having asked a chatbot about their situation, assume the conclusion may be right and the sources are unverified. The paper’s authors could not trust them either without a second agent reading each one. Checking the citation is the part that still needs a lawyer.
The same discipline applies to anything AI-drafted that leaves your firm under a lawyer’s name. A documented human review of every cited authority is what the research says is missing, and it is also what the bar rules already expect.
Source
arXiv, “GANDR: Claim Auditing for Verifiable Legal Answer Generation,” Chen Qian, Yimeng Wang, Yu Chen, Lingfei Wu and Andreas Stathopoulos, submitted September 9, 2026 (preprint, 185-item benchmark) · arxiv.org
03 The structure
Your Maps listing is a composite sketch from 793 sources
A recovered piece of Google’s own Maps software shows the listing you edit is not the entity Google ranks. Your edit is one more piece of evidence.
Search Engine Land published on Tuesday what may be the most detailed public account yet of how Google Maps decides which business to show. Olivier de Segonzac worked from a recovered binary of Google’s internal Geostore system, cross-checked against Maps network traffic and the 2024 leak of Google’s ranking documentation. The headline numbers: 72 ranking signals, 793 data source providers and 446 distinct types of local search intent.
The finding that matters to a firm is about where the listing comes from. “A business doesn’t simply have one source. Its name might come from one provider, its phone number from another, its category from another, and its geometry from somewhere else entirely.”
Google merges those sources into one canonical entity through a process it calls conflation, weighting each provider by how much it trusts it. The listing you see is the interface. The entity Google ranks sits underneath it.
That explains a frustration every managing partner has met. Changing a field in Google Business Profile, he writes, does not mean Google’s canonical record immediately becomes that value: “The edit becomes another piece of evidence entering a system that may already have competing evidence.” If a bar directory, a legacy Yelp page and a former employer’s website all say something different about your address, your own correction is one vote among several, and not always the heaviest.
The ranking signals themselves are less exotic than the industry pretends: reviews, how often people search the name, listing views, direction requests and website clicks, with 25 of the 72 marked deprecated. The question the system is really asking, in the author’s words, is “whether that representation is complete enough for Google to confidently recommend the business.” Everything an AI answer engine says about a local firm draws on that same entity.
What to do about it
Audit the sources, not the profile. List every place your firm’s name, address, phone and category appear (the state bar record, Avvo, Justia, Yelp, Apple Maps, old press coverage) and fix the disagreements at the source, starting with the one Google trusts most, which for a lawyer is the bar record.
Treat a stale listing on a directory you never claimed as a live conflict, not clutter. It is feeding the same entity your prospective client is shown.
Source
Search Engine Land, “Inside Google Maps: 72 ranking signals and the architecture behind local search,” by Olivier de Segonzac, September 8, 2026 · searchengineland.com
04 The one to check today
A stranger’s phone number can appear in your listing’s photos
Fraudsters are uploading AI-generated images carrying their own phone number to other businesses’ Google profiles, and Google’s support desk has not kept up.
Search Engine Roundtable reported this morning on a scam that uses the one part of a Google Business Profile most owners never look at: the photos other people add. A contributor uploads an AI-generated image carrying the target business’s name and a phone number that is not the business’s, and it appears in the listing’s photo gallery beside the owner’s own pictures.
The report came through local-search consultants Naomi Stevens and Sam Sarsten. One account, quoted in full: “A Google Maps contributor has uploaded an AI-generated image to my client’s GBP (roofing company) using the client’s business name but with a phone number that doesn’t belong to the business.” The same contributor, posting as “Jack Edward,” had added more than 150 photos across unrelated U.S. businesses, all carrying the same number.
The mechanism is a hijack without a hijack: nothing in the listing’s data changes, so the owner’s phone field still reads correctly and nothing flags in the dashboard. A prospective client scrolling the photos sees a professional-looking image with a number on it and calls that. Google’s support reply to the report was its generic “we’re not sure of your exact issue” template, and the images had not been removed when the story ran.
Professional firms are exposed in exactly the way the roofer was. A prospective client comparing three lawyers on a phone screen looks at photos more than text, and the listing’s photo gallery is open to any signed-in Google account. It is a stranger taping a business card over the number on your office door, and the door is on Google’s property, not yours.
Do this today
Open your firm’s Google Business Profile, go to Photos, and look at the tab for photos added by others, not just your own uploads. Any image carrying a phone number, email or web address that is not yours gets reported through the flag icon and, separately, through the Business Profile support form, because the generic reply above suggests the first report may not land.
Put the check on a monthly calendar alongside the review sweep. An AI-generated image costs nothing to make, and the report shows one account placing 150 of them.
Source
Search Engine Roundtable, “Google Maps Phone Number in Photo Scam,” by Barry Schwartz, September 10, 2026 · seroundtable.com
05 The money
Google gave a $3,200 ad credit, then took it back
A promotional credit spent in good faith was invalidated a month later, and the advertiser paid the bill with no appeal available.
Google Ads pushes promotional credits at new and returning advertisers: spend a set amount and receive a matching credit. Search Engine Roundtable reported this morning a case where the credit vanished after the money was gone. PPC consultant David Melamed says a $3,200 credit on a client account was invalidated more than a month after the spend had been made against it, leaving the client paying the full amount.
It was not a one-off. “This is the second advertiser account which had the credit invalidated over a month later,” Melamed wrote. In one case the stated reason involved his agency’s billing profile having set up the account; in the other, “I’m not sure why.” His summary of the mechanics: “you obviously can’t crawl back your spend once it’s gone.”
Google’s Ads Liaison, Ginny Marvin, replied publicly: “Thank you for bringing this to our attention, David. I’ve passed this along to the team.” No explanation, no reversal, and as of publication no appeal route exists for an invalidated promotional credit.
Two accounts from one consultant is a small sample, and it is possible both breached a term in the promotion’s fine print. What is not in doubt is that the advertiser found out only after the money was spent.
Small professional firms are the audience for these offers, because the credit is what makes the first month of Google Ads look affordable. A credit that can be withdrawn after the spend is not a discount. It is a coupon the cashier tears up after you have left the store, so budget as though it might not exist.
What to do about it
If your firm or your agency accepted a Google Ads promotional credit this year, open Billing, find the credit line, and confirm it was applied to a paid invoice. Then read the offer’s terms for the conditions that void it, in particular who created the account and which billing profile pays.
Never let an agency’s budget plan count a promotional credit as cash until it shows as applied. Plan the first month at full price and treat the credit as a refund if it arrives.
Source
Search Engine Roundtable, “Bait & Switch: Google Ads Revoking Credits After Advertisers Spend Them,” by Barry Schwartz, September 10, 2026 · seroundtable.com
