People Also Ask Hits 97% AI, EU Results Worst Ever

Estimated reading time: 33 minutes

Two numbers define this week’s search marketing news. People Also Ask answers are now 97% AI-generated, up from about 12% fourteen months ago, and Google has told Reuters that its reworked EU results represent the biggest quality drop in the search engine’s 29-year history. Add a Cloudflare policy change that could block Googlebot on 15 September, Amazon selling ChatGPT inventory through its own DSP, and a content marketing survey sitting at a twelve-year low, and you have a week that touched crawling, ranking, reporting and buying all at once.

DeepMind Puts a Single Model in Charge of Ranking

A Google DeepMind paper proposing one language model to replace the entire two-stage ranking pipeline got a lot of attention this week. I’d like to put a small dent in the excitement.

The proposal is called Autoregressive Ranking. Today’s architecture splits the job: a Dual Encoder turns queries and documents into vectors and retrieves candidates cheaply, then a Cross Encoder, far more expensive to run, reranks what the first stage found. ARR collapses that into a single model that generates the ranked list directly, trained with a method the researchers call SToICaL. In WordNet tests it performed comparably to the expensive Cross Encoder and considerably better than the Dual Encoder.

Then there’s the shopping test. One version of the method lifted the overall ranking score from 95.23 to 97.21 out of 100 while accuracy at picking the correct top result fell from 95.16% to 70.32%. The list got better and the first answer got worse. That comparison covered 310 evaluation examples, and the target rankings were themselves generated using another model’s similarity scores, so what is being measured is agreement with a benchmark rather than agreement with reality.

The paper was first submitted in January, last revised in February, and announces no rollout in Google Search. Candidate documents were supplied to the model in its prompt. Nothing here replaces Google’s web-search infrastructure. File it as a direction of travel rather than a change to plan around.

Google Says Position One to Ten No Longer Maps Cleanly

John Mueller was asked how Search Console counts impressions in the Generative AI report, and his answer went somewhere more interesting than the question.

The mechanics first. Standard impression rules apply to AI Overviews, which means an impression counts when your link appears in the AI Overview on the results page that was served, whether or not the user scrolled to it. Links sitting behind Show More don’t count until someone expands it, so those numbers understate your exposure rather than inflate it. Every link inside an AI Overview is assigned the position of the AI Overview itself, so the average position you are reading is the slot the AIO occupied on the page, not where your link sat among the others in it. The Generative AI report is also a filtered view of data already in your Web search type, so don’t add the two together.

Then Mueller offered this, which is the part to keep:

“Search results pages have a lot of ways for users to interact nowadays, so the old ‘position 1 – 10’ is hard to map, or to make useful for site owners.”
– John Mueller, Search Advocate, Google

Google is tracking these as a block, the way it handles other search features, and Mueller asked openly what site owners would actually find useful instead. Read that as an invitation. Average position has been quietly losing meaning for a couple of years and Google has now said so in public, so if your client reporting still leads with it, this is the week to demote it in favour of impressions, clicks and query-level movement.

Citations Move to the Bottom of AI Mode

Google is showing its Citations Panel at the bottom of AI Mode results more consistently, in a longer format that gives each cited site a title, a description and an image.

For anyone chasing AI visibility that is a trade most would take. A citation card carrying three elements does more for recognition than a compressed inline link, even sitting further down the page. The format was first spotted in July on free listing products and now appears across AI Mode more broadly.

Google Calls Its Own EU Results the Worst Quality Drop in 29 Years

Google has told Reuters that the search results it now serves in the European Union are the worst it has ever shipped.

“The changes mark the largest reduction in quality of service at the world’s most popular internet search engine in its 29-year search history.”
– Google official, quoted by Reuters

The changes came out of the Digital Markets Act, which required Google to stop preferencing its own vertical services. The reworked layout puts one specialised search engine at the top of the page, two more below it with fewer details, and then a carousel of hotels, airlines or restaurants with features such as real-time prices stripped out. Google’s argument is that price comparison sites, Expedia and Booking.com among them, gain prominence over the hotels and airlines in those sectors, which are left with a link, a telephone number and an address. Rankings are still determined by Google’s algorithm. The company tested these changes back in 2024, so the complaint arriving now is not a surprise.

Here is the part UK readers need to hold onto: none of this applies to you. The Digital Markets Act is EU law covering the EU’s member states. The UK left the EU, sits outside the EEA, and its digital markets are regulated separately by the CMA. A UK searcher looking for a hotel in Manchester sees the results Google would have served anyway.

That distinction has teeth for anyone running cross-border campaigns. If you sell into both the UK and the EU from one site, your travel, hospitality or retail traffic may be behaving very differently by market while a blended report shows one flat line. Segment by country before concluding that anything has changed, because the change is real in Paris and absent in Preston.

Cloudflare’s AI Block Starts Catching Googlebot on 15 September

On 15 September, Cloudflare removes the exemption that has kept Googlebot, Bingbot and Applebot outside its AI blocking rules. If your site sits behind Cloudflare with AI training restrictions switched on, and you do nothing, those crawlers can be blocked.

The mechanism matters more than the headline. Cloudflare will assess each crawler against all of its classified uses and apply the strictest restriction it finds. Googlebot handles crawling for Search, AI Overviews and AI Mode from the same agent, so a crawler classified for both search and AI can be blocked even where search crawling is allowed. Google-Extended, the robots.txt control that lets publishers restrict certain Gemini training and grounding uses, is an instruction rather than a separate crawler, and it doesn’t give you an opt-out from those search features either.

The change reaches further than most people assume. It covers:

Sites using the older “Block AI bots” setting
New customers, and new sites added by existing customers
Existing free customers who have not changed their settings before 15 September
Paid customers with training restrictions already enabled, since the expansion of existing blocks is a separate change

Cloudflare is also changing defaults on pages that display advertising: training bots and agents acting for users get blocked by default there, while search stays allowed, subject to the same multi-purpose crawler rules. The reasoning is that an ad implies the publisher expects a human visit, so bots substituting for that visit should face different defaults.

Blocking Googlebot at network level stops it fetching pages for ordinary search results, not only for AI features. That’s an SEO outage dressed up as a privacy setting. Open your Cloudflare Security settings this week and opt out before the deadline if you want Google to keep crawling. The company’s stated aim, pushing mixed-use crawlers to separate search from training, is a reasonable fight to pick. It’s just being picked on your domain.

Rival Search Engines Will See Data Publishers Never Get

Google has published details of its European Search Dataset Licensing Program, and the asymmetry inside it is hard to miss.

Qualifying search providers, including AI chatbots with search functionality, can apply for anonymised ranking, query, click and view data from Google Search. The bar is an average of 50,000 monthly EU users over the previous year, plus either two years offering search in the EU or, for companies less than two years old, more than 50 million euros of investment. Licensing agreements begin on 17 September, with data samples from 16 November.

What competitors receive goes well past Search Console. Under the European Commission’s binding measures they can see the URLs shown in results, their positions, and how users interacted with them: clicks, returns to the results page, hovering, scrolling, swiping and viewing duration. Search Console gives a publisher four headline metrics and omits some queries from its tables entirely.

So a rival search engine may end up knowing more about how people behave on your result than you do.

It isn’t a complete copy of Google Search. Rare queries can be filtered, exact timestamps are removed, only the first query from an AI Mode conversation is included, the data is delayed by at least seven days, and recipients face strict auditing and storage rules. Note the scope too: this is a DMA programme covering search engines operating in the European Economic Area, and the UK is not in the EEA. A UK-based search company would need to qualify on its EU user base. UK publishers, meanwhile, get precisely what they had before, which is Search Console.

Le Monde Says AI Overviews Haven’t Dented Its Audience

Here’s a data point that cuts against the prevailing mood. Le Monde says its audience has not dropped suddenly since AI Overviews launched in France.

CEO Louis Dreyfus, speaking to INfluencia about seven weeks after the French launch, made two arguments. The widely quoted figure of a 33% to 38% traffic loss, which comes from the regulator Arcom and was cited by the press alliance APIG in its complaint to France’s competition authority, refers to Google Search traffic rather than total audience. Search brings around half the visits to major news websites, so a 30% search decline lands as a much smaller overall decline. His second point is structural: Le Monde’s apps generated 36% of group traffic in August, and the company’s view is that AI Overviews mainly affect web traffic.

Le Monde in English, read mainly in the US, England, Scandinavia and India where AI Overviews were already available, grew 7% in the first half of 2026 against the same period in 2025. The group added 24,000 subscribers in July and another 24,000 in August, described as record months by its own standards, though the interview doesn’t tie those numbers to any traffic source.

Caveats belong here. Le Monde didn’t provide a breakdown of its Google Search referrals, and a publisher with a strong app, a paywall and an international edition is not a typical client. The lesson isn’t that AI Overviews are harmless. It’s that a business with more than one distribution channel absorbs the hit differently, which is an argument for building the second channel now rather than after the drop.

Recovery From the August Spam Update Rewards Patience

The August 2026 spam update is confirmed complete, and the recovery advice arriving now is mostly a warning against panic.

Sites that had performed well for years have lost visibility, and a wave of manual spam actions has landed alongside the algorithmic update, which means a fair number of publishers are dealing with both problems at once. Those two need separating before anything gets fixed, because a manual action appears in Search Console and an algorithmic drop does not. Content quality looks like the primary issue, with plenty of speculation pointing at sites using AI to produce content at scale.

Rushed fixes make recovery harder. Stripping pages, reversing recent changes wholesale or rewriting large sections on a hunch destroys the evidence you need to work out what went wrong in the first place.

Retailers have a window here that closes fast. Q4 is close enough that an unresolved visibility problem turns into a cash flow problem rather than a reporting one, and the summer timing meant many teams were running short-staffed when the drop landed.

Relevance and Backlinks Top a Survey of 131 SEO Professionals

Zyppy asked 131 SEO professionals to name Google’s most important ranking factors. Search intent match came first, selected by 57.1%. Backlinks came second at 54.8%.

That second number is the interesting one, because Gary Illyes has said publicly that links are not among Google’s top three signals. The practitioners disagree, and they are specific about which links count: trusted, topically relevant domains, and pages with real visitors. Spammy backlinks ranked among the strongest negative signals in the same survey. Cyrus Shepard, writing up the results, put it plainly, saying the rumour of the death of backlinks has been greatly exaggerated.

Content quality ranked third, and within that category original research and first-party data scored highest of all. AI-generated content was rated favourably where it had substantial human review and added unique value, and negatively where it was published at scale with little added. Meta descriptions sat near the bottom, which will surprise nobody who has run a title-and-meta test recently.

Treat this as a survey of belief rather than a measurement of Google. It tells you what experienced practitioners see working in their accounts, which is genuinely useful, and it tells you nothing directly about the algorithm.

Seventy-Two Ranking Signals Sit Behind a Maps Listing

Local search got its most detailed technical write-up in years, built on a binary exposing a non-public scope of Geostore, the system Google uses to represent geographic entities, cross-referenced with Maps protocols, network traffic, the web index and the 2024 leak.

The headline number is 72 ranking signals. The more useful idea is that the listing you see is not the entity Google holds. Internally a business is a Feature, carrying identity, geometry, source information, websites, business-chain relationships, Knowledge Graph references and ranking information, assembled into a canonical record from several sources. What an owner edits in Google Business Profile is one input to that record, not the record itself.

That reframes a lot of routine local work. If your Business Profile edits sometimes fail to change anything, it’s because you are editing one source feeding a canonical entity that other sources also feed. Consistency across those sources does more than repetition within any one of them. The piece also argues that Maps is turning into a conversational product, which is where the local and AI stories meet.

Content Marketing Confidence Hits a Twelve-Year Low

Just 14% of marketers say their blog delivers strong results, according to Orbit Media’s annual blogging survey. That’s six percentage points below the previous twelve-year low, and close to half the 26% reported in 2022.

I’ll admit this one landed awkwardly. A fair amount of what our industry has sold for a decade rests on the premise that publishing consistently produces compounding returns, and the people doing the publishing are increasingly saying that it doesn’t.

The survey’s explanation is not the one you would expect. AI adoption reached 92.4% and showed no relationship with stronger results in either direction. What correlated with the decline was the abandonment of practices historically linked to success: influencer collaboration, original research, keyword research, paid content promotion, formal human editing and consistent use of analytics. Marketers who did keyword research were more likely to report strong results, which is a quietly satisfying finding for anyone who has been told that SEO stopped mattering.

AI did cut production time. The average blog post now takes 3 hours and 20 minutes to write, down from more than 4 hours in 2022, saving a typical marketer roughly 50 hours of writing a year while reported performance keeps sliding. Ann Handley of MarketingProfs called the findings a wake-up call and said marketers need to be much more discerning about which efforts they remove and which they keep. Speed replaced the parts of the process that made the output good, and this survey is what that trade looks like a year later.

Amazon Starts Selling ChatGPT Inventory Through Its DSP

Amazon is piloting the sale of ChatGPT advertising through its demand-side platform, connecting one of the largest advertising businesses on the internet to the newest conversational channel.

Advertisers can buy on a cost-per-click or CPM basis, and Amazon is offering product feed ads that build assets automatically from a product catalogue. The units appear as text and image placements beneath organic ChatGPT responses, carrying sponsored labels. Testers receive aggregated performance data: impressions, clicks, cost per result, CPM and CPC.

The split of control is the detail to hold onto. Amazon provides the buying and campaign management layer while OpenAI keeps control of ad delivery and placement inside ChatGPT. Buying through Amazon does not mean Amazon decides where you appear.

What makes this more than a plumbing story is Amazon’s shopping and streaming data arriving next to conversational intent. That combination hasn’t existed before, and for retail advertisers it’s the first version of ChatGPT advertising that connects to an audience graph they already work with.

ChatGPT Ads Go Global While Rival AI Tools Get Shut Out

Two OpenAI advertising stories landed this week, and they read differently side by side.

The first is expansion. ChatGPT Ads are rolling out to selected countries across Europe, India, the Middle East and North Africa, with a batch of platform upgrades alongside: custom audiences that can be edited without rebuilding from scratch, mixed identifier types in a single request, audiences above five million members, looser exclusion-audience restrictions, Google Advertising ID support, and a Measurement Pixel and Conversions API that now accept additional hashed customer information including phone numbers, names, regions and postal codes. That is a performance advertising stack taking shape, not a novelty placement.

The second is restriction. OpenAI has reportedly told advertising partners it will no longer approve campaigns promoting standalone image and audio generation products, a change communicated privately rather than announced. Adobe is among those affected, having promoted Acrobat Studio and its Firefly image generator during the original pilot. Video generation tools can still advertise, according to the report.

Platforms blocking direct competitors is ordinary behaviour and every large ad business does some version of it. The thing to watch is that OpenAI is expanding its own image and voice capabilities at the same time as closing the door on advertisers who sell those things, and setting that policy in private. Anyone planning ChatGPT spend should check their product category against a rulebook that can move without notice.

One scoping note for UK readers: the report describes the rollout as reaching selected countries across Europe, which is neither the EU nor a guarantee that the UK is in the first wave. Check availability in the account rather than assuming, because “Europe” in US trade coverage rarely means what a UK reader takes it to mean.

What 97,131 ChatGPT Ad Impressions Actually Bought

SE Ranking ran its own ChatGPT Ads test and published the numbers, which is more useful than another explainer.

The test ran in June 2026 across the US, Canada, Australia and New Zealand:

Three campaigns, eight ad groups, 48 ads
97,131 impressions and 1,263 clicks
An average CTR of 1.30%
Almost no sign-ups, and the test was stopped before the planned budget was spent

Delivery was lopsided. Two of the eight ad groups produced around half of all impressions, because volume depends on how many conversations match each ad group’s targeting rather than on a query you bid against. Targeting here means conversation context plus keyword-style phrases plus negatives, so a Google Ads structure copied across is unlikely to work.

For context, ChatGPT Ads launched in the US on 9 February 2026 and now runs in 40 markets, shown to logged-in adults on the Free and Go plans while Plus, Pro, Business, Enterprise and Edu accounts stay ad-free. Three campaign objectives are available: Reach, Clicks and Conversions.

A 1.30% CTR producing no sign-ups is the kind of result most companies bury rather than publish, so credit to SE Ranking for putting it out. The honest reading is that traffic from conversational ads is not yet demand, and anyone budgeting for this channel should plan a test they can afford to lose.

Google Ads Is Revoking Credits After Advertisers Spend Them

PPC consultant David Melamed has now had it happen twice: Google Ads issues a promotional credit, the advertiser spends against it, and more than a month later the credit is invalidated.

In one account the sum was $3,200, spent on the understanding that it amounted to free advertising. No appeal route for an invalidated credit has been identified, and the spend obviously can’t be recovered. Melamed’s account is that one affected advertiser was a brand new client whose credit was invalidated because his MCC billing profile had set up the account initially. Ginny Marvin, the Google Ads Liaison, replied publicly that she had passed it to the team.

His wider point is the one agencies should sit with. Advertisers holding a credit bid more aggressively than they otherwise would, which lifts auction prices for every other advertiser competing for those impressions, including the ones who never saw a credit.

Practical response: treat promotional credits as provisional until they have survived a couple of billing cycles, put that in writing to the client before the spend happens, and screenshot the credit terms at the point of acceptance. That’s a depressing amount of defensive admin for a promotion. Right now it’s also the sensible amount.

Exact and Phrase Match Find a Way Into AI Mode

Google has begun a small experiment letting traditional Search campaigns serve text ads in AI Mode using exact and phrase match keywords.

Until now the route into AI Mode ran through AI Max or Performance Max, which meant giving up keyword control to get the placement. Ginny Marvin confirmed the test, first spotted by AdSquire founder Anthony Higman across multiple campaigns. Eligibility is limited to queries where Google identifies explicit and direct user intent, so this is not a general opening.

That limit is the interesting part. Google is conceding that some AI Mode queries are ordinary transactional searches wearing a conversational coat, and for those a tightly matched keyword campaign is the right instrument. AI Max and Performance Max stay positioned for the messier, longer, more exploratory queries where Google wants its automation making the decisions.

For advertisers who resisted shifting budget into automated products purely to reach AI Mode, this is a reason to hold that position a little longer.

AI Mode Fills the Product Block With Ads

Search Engine Watch tested ecommerce queries in AI Mode this week and found something new. On a search for headphones, every product shown inside the answer was an ad. Not one organic product listing appeared.

Previously AI Mode surfaced organic product results within its answer, with ads kept separately at the bottom. If the version spotted here spreads, the product block inside an AI answer becomes a paid surface and organic merchandising loses its place in it altogether.

For retail clients that collapses the usual planning separation between SEO and paid shopping. If presence in the AI answer for a commercial query requires a feed and a budget, the organic team’s ceiling on those queries gets set by somebody else’s media plan. Nobody should redraw a strategy around a single observed test, but this one earns a monitoring line rather than a shrug.

Target CPA Stops Beating Its Own Target

Google’s change to target bidding has a simple effect and a long memory.

Previously, Target CPA and Target ROAS behaved like efficiency safeguards. A campaign with a $10 Target CPA that could deliver conversions at $5 would carry on delivering them at $5. Under the new approach the target is taken literally: set it at $10 and Google seeks conversions around $10 rather than trying to beat it.

You get predictability in exchange. Forecasting what happens when a budget increases becomes more reliable, because the system holds a defined efficiency level instead of finding whatever it can. What you lose is the free upside on accounts that were quietly outperforming, and those accounts will watch that advantage shrink.

Reva Minkoff’s point in the SMX Now session is that this is where Target CPA started. Around 2015 and 2016 Google described it as a strategy that set bids so the average cost per conversion would equal the advertiser’s target. The behaviour everyone got used to came later. If an account has been beating its targets, reset the target to the number you actually want rather than the one that used to produce a bonus.

AI Overviews Are Quietly Moving Shopping Ad Numbers

Shopping ad click-through rates are rising while impressions fall, and Mike Ryan, head of ecommerce at Smarter Ecommerce, thinks AI Overviews may explain the pattern.

Across thousands of Shopping and Performance Max campaigns spanning hundreds of advertiser accounts, Ryan found clicks relatively flat or slightly down while impressions had fallen more noticeably. Separate ecommerce data from Optmyzr showed CTR up 17% year over year, close to the roughly 20% increase he was seeing in Smarter Ecommerce’s own benchmark data. A rising CTR here is arithmetic rather than improvement, because the denominator shrank.

His hypothesis is that Google may be more likely to serve AI Overviews on queries with a lower predicted likelihood of producing an ad click, while preserving Shopping placements on the queries most likely to convert. Publishers get what Ryan calls the crocodile effect, gaining impressions and losing clicks. Shopping ads are seeing the mirror image of it.

The reporting consequence is immediate. A CTR improvement in a Q4 board pack that is really an impression decline will be read as a win by everyone who sees it. Put impression volume, clicks and total traffic next to CTR, or the number tells a story that isn’t happening.

Merchant Center Adds AI Search Intent, Terms and Attributes

The Merchant Center AI Performance Insights report, which launched in July and began expanding to more regions this month, has gained three sections: AI Search intent, AI Search terms and AI attributes.

Spotted by Brodie Clark, the additions show how your products match customer AI searches, recommend popular terms to include in product descriptions for products appearing in AI searches, and flag missing attributes on those products. Google also made a number of changes to the report’s help documentation at the same time.

This is the first reporting Google has offered that ties a specific product feed to AI search visibility, which makes it the first place an ecommerce team can look instead of guess.

Data Manager Spreads and Meridian GeoX Goes Global

Google put out a slate of measurement releases this week, and the through-line is first-party data becoming the thing you get scored on.

Data Manager is expanding beyond Google Ads into Google Analytics and Display & Video 360, and its API is becoming universal through the adoption of the IAB Tech Lab’s ECAPI standard. In Google Ads, a new Data Strength Uplift Metric quantifies the value generated by an advertiser’s first-party data setup, which puts the business case for a data project in the same interface as the spend it has to justify.

Meridian, Google’s open-source marketing mix modelling platform, is gaining agentic features that audit data quality, troubleshoot errors and offer guidance while models are being built, lowering the technical bar for MMM considerably. It will also accept brand signals such as Branded Google Query Volume, letting advertisers connect upper-funnel spend on television and out-of-home to brand demand and eventually to sales. Meridian GeoX, which runs causal geo-experiments to measure incrementality and feeds the results back into the models, is now available globally after a beta earlier this year.

For agencies, GeoX going global is the practical one. Incrementality testing has always been the gap between what a client believes about a channel and what that channel actually does, and the tooling for it is now open-source and supported.

Chatbots Talked 57.5% of Shoppers Out of a Purchase

Exploding Topics surveyed more than 2,300 people in the US about how they use AI in their purchasing journeys, and the standout finding is a negative one. More than half of AI users, 57.5% of them, decided not to buy something based on what a chatbot told them.

Almost two-thirds have at least partially replaced product-related Google searches with AI conversations. People are researching local businesses and large ones alike, and their AI usage is increasing over time. So the funnel is real, and there’s a gatekeeper sitting in the middle of it forming opinions about your product before you get a look in.

More than four in ten people say they actively dislike AI chatbot ads, which complicates the obvious response of buying your way back into the conversation.

What this argues for is unglamorous: make sure the sources a chatbot reads about you are accurate and current. Reviews, specification pages, comparison content, third-party write-ups. When a model talks a customer out of a purchase it does so using material it found somewhere, and most of that material is fixable by someone on your team this quarter.

Mullenweg Is Out at Automattic and Still Running WordPress

Matt Mullenweg has been removed as CEO of Automattic and placed on paid leave, with CFO Mark Davies stepping in as interim chief executive. Mullenweg says he received the board resolution 50 minutes before the meeting began, and that his repeated requests for time to have it reviewed by independent legal counsel were denied.

Automattic’s official statement confirms the leave and expresses full confidence in Davies. Separately, WordPress Executive Director Mary Hubbard moved quickly to state that WordPress leadership and direction remain unchanged and that Mullenweg continues to lead the project. So the chief executive of the company is out while the leader of the open-source project is not, which is the governance knot WordPress has never untied.

Context shapes how this reads. It follows sustained negative sentiment over his dispute with WP Engine, a lawsuit against Automattic arising from those actions, and a motion for sanctions alleging failure to preserve evidence, including reports that a laptop and a phone were unaccounted for.

For the sites already running WordPress, nothing changes this week. But anyone making a five-year platform recommendation for a client now has a live governance question to put in the risk column, and that is new.

Strategic Direction: Where Search Marketing Is Heading

Three things happened this week that belong in the same sentence. People Also Ask went almost fully AI, Google told a news agency that its EU results are the worst it has shipped in 29 years, and Cloudflare prepared to pull search crawlers into its AI blocking rules. Each on its own is a story. Together they describe a search environment where the answer layer, the regulatory layer and the access layer are all being rewritten by different parties at the same time, and none of those parties are coordinating with each other.

The measurement thread running underneath is just as loud. Mueller conceded that position 1 to 10 no longer maps usefully. Shopping CTR is rising because impressions are falling. Rival search engines will soon see interaction data that publishers cannot get about their own results. Google is shipping a metric that scores the strength of your first-party data. Every one of those points the same way: the numbers that described search for twenty years are becoming less informative, and the teams that replace them first will spend the next year explaining performance more accurately than their competitors can.

For paid media specifically, the conversational channel stopped being theoretical this week. Amazon is selling ChatGPT inventory through its DSP, OpenAI is building a genuine ad stack while setting policy privately, exact match found a door into AI Mode, and SE Ranking published a test in which 97,131 impressions produced almost no sign-ups. Both of those facts are true at once. The channel is becoming properly buyable, and it has not yet been shown to produce demand. Budget accordingly, which means testing with money you can afford to write off and measuring something considerably further down the funnel than a click.

Key Takeaways

PAA answers reached 97% AI-generated in the first week of September, up from about 12% fourteen months ago. Reclassify question-led content as a visibility surface rather than a traffic source before your reporting does it for you
The EU’s DMA changes do not apply in the UK, which sits outside both the EU and the EEA. Segment cross-border reporting by country before concluding that a travel or retail account has changed behaviour
Cloudflare removes its search-crawler exemption on 15 September. If your site has AI training restrictions enabled, check the Security settings this week or risk blocking Googlebot from ordinary search crawling, not just AI features
Mueller has said in public that position 1 to 10 is hard to map usefully. Demote average position in client reporting and lead with impressions, clicks and query-level movement instead
Rising Shopping ad CTR is being produced by falling impressions rather than better ads. Report impression volume alongside CTR or the Q4 numbers will mislead the people reading them
Target CPA and Target ROAS now aim at the target instead of beating it. Reset targets on accounts that were quietly outperforming, because the free upside is going away
SE Ranking’s published ChatGPT Ads test produced 97,131 impressions, 1,263 clicks and almost no sign-ups. Treat conversational ad budget as a test you can afford to lose, with conversion measurement in place before launch

Frequently Asked Questions

Do the EU’s DMA search changes affect UK websites?

No. The Digital Markets Act is EU law, and the UK sits outside both the EU and the EEA, with its own digital markets regime run by the CMA. UK searchers see results unaffected by the changes Google described to Reuters. If you sell into both territories from one site, segment reporting by country, because the same pages can behave very differently in Paris and in Preston.

What do I need to do before Cloudflare’s 15 September change?

Open your Cloudflare Security settings and check whether AI training restrictions are enabled, including the older “Block AI bots” option. From 15 September those rules can apply to Googlebot, Bingbot and Applebot, and blocking Googlebot at network level stops it crawling for ordinary search results rather than only for AI features. You can opt out of the change before the deadline.

Is People Also Ask still a sensible target now the answers are AI-generated?

Yes, but measure it differently. A PAA appearance now works like a citation inside an AI-generated answer rather than a route to a click, so judge it on visibility and brand exposure instead of sessions. Content that answers a question precisely and adds information the other results lack is still what gets pulled in.

Should we be buying ChatGPT ads yet?

Only with budget you can afford to lose, and only with conversion measurement running before launch. SE Ranking’s published test generated 97,131 impressions and 1,263 clicks but almost no sign-ups, and delivery was heavily uneven across ad groups. The channel is becoming properly buyable now Amazon’s DSP is selling it, which is a separate question from whether it converts.

Why is our Shopping ad CTR rising while performance feels flat?

Because CTR rises when impressions fall faster than clicks do, and analysis from Smarter Ecommerce suggests AI Overviews may be reducing Shopping ad impressions on lower-intent queries while leaving high-intent ones alone. Put impressions, clicks and total traffic next to CTR in your reporting so the ratio gets read in context.

Conclusions

A week that rewrote the answer layer, the access layer and the measurement layer at once leaves most teams with the same immediate job: work out which of your numbers still mean what you think they mean. Average position is fading, Shopping CTR is being flattered by falling impressions, PAA has stopped being a traffic source, and a first-party data score is about to appear next to your ad spend. None of that calls for a strategy rewrite. It calls for fixing reporting before Q4, because every one of those metrics is about to be read by somebody making a budget decision.

Two things need doing this week rather than this quarter. Check your Cloudflare settings before 15 September, and segment any cross-border account by country so the EU changes don’t hide inside a blended line.

Need help adapting your search strategy for the AI era? Contact the Anicca team for expert SEO and PPC guidance.

Table of Contents

AI Answers and How They Are Counted

Regulation, Access and Publishers

SEO Strategy and Rankings

Paid Media and AI Advertising

Measurement and Ecommerce

Industry

Wrap-Up