AI News Roundup — September 13, 2026

Google Finished a Spam Update in 64 Hours, and It Was Aimed at Volume

Google's August 2026 spam update started rolling out on August 18 and was finished on August 21 — roughly two days and sixteen hours, global, across every language. It was the third spam update of the year, following rollouts in March and June, and Google announced no new policy categories alongside it. The existing spam policies are still the document to check yourself against.

Two details are worth sitting with. The first is the speed. A 64-hour global rollout is not a system being carefully eased into place; it is a system confident enough in its classification to apply it at once. The second is the target. The pattern reported across affected sites points at scaled content abuse — pages produced in volume to cover keyword combinations rather than to answer anything, whether a machine or a content farm produced them. Note carefully what that is and is not. It is not a penalty on using AI to write. Google's stated position remains that the method of production is not the issue; the issue is publishing at a scale and a thinness that exists to occupy search results.

For a local business this is less abstract than it sounds, because the single most common thing a marketing vendor sells a local business is exactly the pattern this update punishes: a page for every service crossed with every town in a fifty-mile radius, four hundred URLs, identical but for the place name. That has been a bad idea for years and it is now an actively dangerous one. We have made this argument in the contractor SEO guide repeatedly; the difference is that the cost used to be wasted money and is now a sitewide risk.

The Test That Separates a Safe Page From a Risky One

Open any page on your site and ask whether it contains a single fact that came from operating your business — a constraint you have run into, a price range you actually charge, a building or a route or a piece of equipment you know by name, a thing you will not do and why. If the answer is no on a hundred pages, you do not have a content library, you have a surface area. If the answer is yes on twelve pages, those twelve are worth more than the hundred and they are also the only ones an assistant will ever quote. If your pages are not appearing at all and you are not sure whether that is suppression or something simpler, our guide to a website not showing up on Google covers the ten-minute version of that check.

Two Thirds of Searches End Without a Click — and AI Mode Is Not Why

This is the story most worth reading carefully, because the number everyone is quoting and the conclusion everyone is drawing from it do not actually connect.

SparkToro's 2026 clickstream analysis found that about 68 percent of Google searches in the first four months of the year ended without a click to any external site, and that the share of searches producing at least one click fell by roughly 9.5 percentage points between 2024 and 2026. Those are third-party panel estimates rather than audited figures, so hold the decimals loosely. The trend is not in dispute.

Here is the part that got lost. The same analysis found that AI Mode — Google's conversational search surface, the one every vendor is currently building a service around — accounted for roughly one third of one percent of all Google searches in that window. Both numbers are true at once, and together they make an arithmetic point that is hard to argue with: a feature used in 0.34 percent of searches cannot be responsible for a 68 percent outcome. It is not large enough to move the total no matter how zero-click its own behavior is, and its own behavior is extremely zero-click — somewhere between 92 and 94 percent of AI Mode sessions end without an outbound visit.

So what did take the clicks? Things that have been accumulating for ten years and that nobody sells a product to fix: AI Overviews sitting above ordinary results on a large share of queries, the local map pack answering the question in place, featured snippets, knowledge panels, the weather and sports and unit-conversion boxes, and Google's own properties appearing in its own results. The search results page gradually became a destination instead of a doorway. AI Mode is a real and fast-growing surface — Google said at I/O this year that it had passed a billion monthly users with query volume more than doubling each quarter — and it will matter more next year than it does today. It is simply not what happened to your traffic between 2024 and now.

Why the Correction Changes What You Should Buy

If you believe AI Mode took your clicks, the obvious purchase is an AI-search visibility program, and there is a growing industry happy to sell you one. If you understand that a decade of answer-in-place features took your clicks, the obvious purchase is different: be the answer that gets shown, and make sure the queries where a click still happens — the considered, high-value, I-am-about-to-spend-money ones — land on a page that converts. Those queries still click, because nobody chooses a contractor or a clinic from a summary box.

That does not make answer engine optimization a waste — it is the correct long-term bet and we build it for clients — but it should be bought as an investment in where search is going rather than as an emergency repair for a wound something else inflicted. The measurement problem underneath all of this is the subject of our longer piece on how much demand is now invisible to analytics, and what assistants actually draw from when they name a business is in what AI assistants actually cite.

The Credential Became the Product

Anthropic published its September threat intelligence report on September 10, covering misuse observed between December 2025 and August 2026 across seven categories. Most coverage focused on the state-sponsored espionage cases, which are genuinely alarming and have nothing to do with running a plumbing company. Two findings in it are aimed squarely at ordinary businesses.

The first is fraudulent AI resellers: operations advertising cheap access to frontier models and collecting customer credentials in the process. Think about who inside a small business takes that offer. It is not the owner. It is the capable person doing the most with AI on the least budget — and the credentials they hand over are frequently attached to the same accounts that run everything else.

The second is the industrialization of credential harvesting. The report documents an operator who downloaded roughly 1.8 million Android applications and decompiled them to extract secrets developers had left embedded in the code, and a case where a single stolen developer token became full administrative control of a victim's cloud environment in about three hours. The framing that stayed with me is that a stolen API key is now valuable three separate ways: it can be resold, it can be spent as compute on somebody else's bill, and it makes the attacker's traffic look like yours. That third one is the reason this matters even if you think you have nothing worth stealing.

None of the responses are sophisticated. Buy AI access from the vendor or a reseller you can verify by name. Keep a written list of which keys exist and who holds them. Rotate anything that has ever been pasted into a shared doc, a ticket, or a chat thread. And if a contractor built something for you that calls an AI service, ask them where the key lives — if the answer is that it is in the website's front-end code, it is already public. This is the same ownership question we work through in do I own my website, applied to a newer asset.

The One Item You Can Act On: It Got Cheaper Again

Anthropic shipped Claude Fable 5.1 on September 1, holding its headline token rates steady while cutting cache-read pricing by about 75 percent. The company said typical workloads end up costing roughly a quarter less overall, with heavily agentic workloads closer to 45 percent less — vendor-stated figures, so directional rather than audited.

Cache reads are the boring part of the bill and that is exactly why the cut matters. Any system that does real work — reads your documents, keeps context across a long task, answers from a body of knowledge about your business — re-reads the same material constantly. That re-reading was quietly the dominant line item in agentic workloads. Making it cheap does not make AI smarter; it makes the long-running, context-heavy systems that were marginal on cost into systems that are comfortably affordable.

The wrong conclusion is to wait for it to get cheaper still. The token cost was never the expensive part of a small-business AI project — the expensive part is deciding what the system should do and writing down what is true about your business, and that does not fall in price. It is also worth noticing what this story does to last month's. Frontier labs are becoming ordinary, margin-conscious companies, and the same month one of them moves toward a public listing it cuts a price by three quarters. Both things are true: the era of capability sold below cost is ending, and unit prices keep falling because the underlying efficiency keeps improving. Plan for portability rather than for a price shock.

The Through-Line: You Are Downstream of Systems You Do Not Operate

Read the four together and something uncomfortable shows up. Three of them are things that happened to businesses, not things businesses did. An algorithm re-sorted the web over a long weekend with no notice, no appeal, and no explanation of which of your pages it disliked. The click-through environment moved under everyone who depends on search traffic, for reasons that were never announced because they were never a single decision. Criminals pointed automated systems at credentials belonging to companies that will never know they were scanned. Only the fourth — capability getting cheaper — is a lever you can pull.

That is not a reason for fatalism, but it should change where you put effort. Anything whose value depends entirely on a system you do not control is a rented position, and rented positions get repriced without your consent. What survives an August weekend is the stuff nobody else can produce: a website you own outright, pages containing operating knowledge that only your business possesses, customers who know your name and can reach you without a search engine standing in between. Every one of those is slower to build than the alternative. All three are still there on Monday.

What to Actually Do in the Next Two Weeks

Four things, in order of how cheap they are. Pull the four weeks before and after August 18 in Search Console and look for a section-shaped loss rather than a general one — that tells you whether the spam update is your problem or somebody else's. Count your pages into the two piles described above, the ones with operating knowledge in them and the ones without; you are looking for the ratio, not a decision. Write down every AI service credential your business has issued and who holds it. And check what your highest-value page looks like to somebody who arrived ready to buy, because that is where the clicks that still exist are going — our post on traffic without leads is the diagnostic for that, and mobile conversion is where that failure usually hides.

If the outcome of that audit is that the site itself is the constraint, that is AI website building work, and the honest numbers for it are in our small business website cost guide. If the outcome is that you need pages worth citing rather than more pages, that is SEO and AEO together. We build both with Claude Code, which is how a real content structure ships in days rather than months — and why we can write twelve pages that know something instead of four hundred that do not.

What We Are Watching Next

Three things. Whether a fourth spam update lands before the holiday retail window, which would be unusual timing and would catch a lot of businesses mid-campaign. Whether AI Mode's share of total searches moves off a fraction of a percent, because the day it does, the analysis above stops being true and the AI-visibility argument becomes urgent rather than strategic. And whether the pattern in the threat report — automated operations run against small targets because the marginal cost of trying is now near zero — starts producing incidents at businesses small enough that nobody ever wrote a security policy. That last one is the story we expect to be writing about by winter.