01 The deadline
OpenAI will rewrite your ChatGPT ads from October 30
Your advertisement can go live in words you never wrote, on a campaign you set up months ago, from October 30.
The change covers existing self-serve campaigns as well as new ones, and declining it has to be done before the date rather than after.
Barry Schwartz at Search Engine Roundtable published the advertiser email on 20 October. In OpenAI’s words, “Starting Friday, October 30, 2026, we’ll enable creative enhancements for your new and existing self-serve campaigns.” The word doing the work there is existing. A campaign built, checked and approved months ago is included by default.
What the setting changes is not cosmetic. OpenAI says it may create new or updated headlines and descriptions, translate the advertisement text into another language, resize images, and send people to a different page on your website by optimizing the destination link. Those variations may be prepared in advance or created in real time based on the conversation.
Then the sentence that decides whether this is a convenience or a problem. “When enabled, these ads can run automatically without requiring you to review or approve each version.” For a business with nobody to answer to that is time saved, and for most of the businesses reading this approval is not paperwork but the step a license depends on.
Three examples, and none of them is a law firm. A charter broker operating under 14 CFR Part 295 may never be presented as the carrier, so a tightened headline that turns “arrange your flight” into “fly with us” has made a claim the rule forbids.
A builder advertising in California has to carry the license number in every advertisement, and a reworded creative is exactly where a number goes missing. A specialty veterinary hospital may claim round-the-clock care only where that is operationally true, which is the kind of reassuring phrase an optimizer is built to add.
What to do about it
Open OpenAI’s Ads Manager before 30 October and set the preference you actually want at account level, then check it again at campaign level, because OpenAI says campaign-specific settings take priority over account defaults. Existing opt-outs are preserved, so if you already declined this, confirm it rather than assume it. If you decide to leave enhancements on, the person who signs off on your advertising claims needs to know that what goes out is no longer the wording they approved.
Source Search Engine Roundtable, Barry Schwartz, “OpenAI Auto Enabling Creative Enhancements In ChatGPT Ads,” 20 October 2026, quoting the advertiser email OpenAI sent · seroundtable.com/openai-chatgpt-ads-enabling-creative-enhancements-42344.html
02 The research
Researchers tested 13 AI models and blamed the search step
When a customer’s question needs two facts about you, the step that goes and fetches pages usually finds one and stops.
The models read what they were given perfectly well; what reached them was incomplete, and scoring only the final answer hid the gap.
Debopriyo Banerjee and five co-authors published the study on arXiv on 18 October, the preprint site where most artificial-intelligence research appears before a journal sees it. They tested ten proprietary frontier models and three open-weight ones, on documents running from 10,000 to a million characters.
Their finding, in their own words: “similarity retrieval surfaces any relevant chunk far more reliably than all required chunks for multi-fact queries.” In plain terms, the fetching step is good at finding one relevant piece of a page and unreliable at finding every piece a question actually needs.
Two consequences land on any business hoping to be recommended. Cutting documents into pieces destroys the document-level context that a question about one named business depends on. And because most testing scores only the final answer, a reply can look properly sourced while its citations are incomplete.
This is the part worth sitting with, because customers do not ask single-fact questions. “Is there an emergency veterinary hospital near me with an overnight intensive care unit that takes my insurance” is three facts at once, and if those three facts live on three separate pages of your website, the fetching step is likely to return one of them and move on.
What to do about it
Put the facts a customer combines in a single question onto a single page. A specialty veterinary hospital that states its hours, its intensive-care capability and the insurers it accepts in one place is more findable than one with a separate page for each. The same holds for an insurance broker’s licensed states alongside the lines they write, or a surgical practice’s subspecialty alongside its hospital affiliations, which means consolidating what you already have rather than writing more.
Source arXiv, Debopriyo Banerjee, Owais Aijaz, Ahmer Tabassum, Zhuohan Xie, Preslav Nakov and Veselin Stoyanov, “The Bottleneck Is Retrieval, Not Reading,” 18 October 2026 · arxiv.org/abs/2610.26951 (a preprint, not yet peer reviewed; ten proprietary and three open-weight models, context lengths of 10,000 to one million characters, on the MultiHop-RAG and QASPER-MP benchmarks)
03 What people actually ask
Seven in ten legal chatbot conversations wanted personal advice
People are not asking an assistant what the rules say. They are asking it what they personally should do next.
The sample is one public archive of real conversations rather than a representative survey, which makes it a direction worth knowing rather than a measurement of your market.
Amirreza Mohseni and Gerasimos Spanakis published their study on arXiv on 18 October. They built a classification scheme for legal questions and applied it to 2,122 real conversations drawn from WildChat, a public archive of chatbot exchanges, checking their labels against a human-reviewed set of 200.
The figure they report: “In this enriched sample, 71.1% of conversations containing legal content seek personalized or practical guidance.” The most common request was advice about rights and options, and 25.3% of the conversations seeking guidance covered more than one legal topic at once.
Be careful with that number. The sample was deliberately enriched to contain legal content, and WildChat records people who chose to use one set of chatbots, so it describes chatbot users rather than the general public. The authors have published their data, models and code, which is more than most studies in this field do.
The pattern is not confined to law. Someone who brings a layered personal problem to an assistant and asks what to do is doing the same thing whether the subject is a severance offer, an unexpected tax notice, an estate or a diagnosis. The question that used to arrive as a phone call is arriving as a conversation with software, and whoever gets named inside it is the shortlist.
What to do about it
Write the page that answers the question a worried person actually types, in their words, in one place. An accountant whose site sets out what to do in the week a notice arrives, or an admissions consultant whose site addresses a deferred application, is answering a real request rather than describing a service. Keep the review step exactly where it is: every industry we work in needs a named human to stand behind advice, and nothing in this study changes that.
Source arXiv, Amirreza Mohseni and Gerasimos Spanakis, “Legal Guidance in the Wild: How Users Seek Legal Help in Real-World LLM Conversations,” 18 October 2026 · arxiv.org/abs/2610.25159 (a preprint, not yet peer reviewed; 2,122 conversations from the public WildChat archive, deliberately enriched for legal content, with a 200-conversation human-reviewed set)
04 The platform
Google will check financial advertisers against a bank registry
Google is moving from asking who you are to asking whether an outside registry says you are allowed to advertise.
The program reaches the United States and Canada in early 2027, and Google has not said which advertisers beyond banks it covers.
Keerat Sharma, Google’s vice president for advertisement privacy and safety, wrote on 20 October that the Financial Services Verification program comes to the United States and Canada in early 2027. In the United States, Google will work with the American Bankers Association to use their dedicated registry to verify authorized banks before ads can run.
The standard being set is a different kind from the one before it. Today’s checks ask an advertiser to prove identity, and Google says more than 97% of the advertisements people see in the United States and Canada already come from identity-verified advertisers, measured over the year to 31 August 2026. The new program instead asks an advertiser to demonstrate they are authorized by a regulator or a recognized registry.
What Google has not said is who else is in scope. The announcement names banks and the two bankers’ associations, and it does not say whether a registered investment advisor, an independent insurance broker or a tax resolution business will face the same gate. “Certain financial ads” is as precise as the policy wording gets, and Google says advertisers will be told if and when they need to complete the process.
One figure in the same announcement deserves a flag rather than a repetition. Google says that in 2025 its systems caught over 99% of policy-violating advertisements before they served, which is Google measuring Google with no outside audit. We are not going to pass that number on as a fact, and the same test should be applied to any visibility figure put in front of you, including ours.
What to do about it
If you advertise in anything a reasonable person would call a financial category, find out now which registry would evidence your authorization, and check that your entry in it is current and correct. A stale registry record is the sort of thing that takes months to repair and stops an advertisement in the week you need it running. Wealth advisors, independent insurance brokers and tax professionals are better off assuming the scope may widen than waiting to be told.
Source Google, Keerat Sharma, “Expanding our defenses against financial scam ads,” 20 October 2026 · blog.google/products/ads-commerce/financial-services-verification-united-states-canada (Google’s own figures, including the 97% identity-verified share measured 1 September 2025 to 31 August 2026; no external audit is cited) · reported by Search Engine Roundtable
05 The money
Google says editing a campaign does not restart learning
If you have avoided touching a campaign in case it reset, Google has now written down that the fear was misplaced.
The rewritten help page also puts a ceiling on the learning period that advertisers have argued about for years.
Barry Schwartz reported on 21 October that Google has rewritten its help page on Smart Bidding, the setting that hands each individual bid to Google to choose. Two statements in it change how an account should be handled. The first is a ceiling: the learning status never lasts more than 7 days.
The second is the one that has been costing money. Google says that changes to budgets, targets, creatives or targeting settings never remove historical data or trigger learning status. A widespread belief in advertising has held the opposite, and it has left underperforming campaigns running untouched for fear of resetting them.
The practices Google now recommends are specific. Keep daily budgets high enough that the system has room to work, and consolidate ad groups and campaigns so that conversion data pools into larger sets.
Optimize toward a single primary conversion goal, and avoid several target changes inside one conversion cycle without a clear business reason.
That consolidation advice cuts against the way many service businesses build an account. A real estate brokerage running a separate campaign for each neighborhood, or an admissions consultant running one per program, is splitting the conversion data the bidding depends on. Fewer and larger is now Google’s own recommendation rather than an agency opinion.
What to do about it
Open the campaign you have been leaving alone and make the change you have been putting off, because the reason you were waiting is one Google has now contradicted in writing. Then look at whether your account is divided into more campaigns than your monthly conversion count can support. That division is now the thing holding your results down rather than the thing protecting them.
Source Search Engine Roundtable, Barry Schwartz, “Google Ads Updates Smart Bidding Learning Doc With 7-Day Limit & Best Practices,” 21 October 2026, on Google’s rewritten “About Smart Bidding learning” help page · seroundtable.com/google-ads-updates-smart-bidding-learning-doc-42345.html
