Digest

62% of Your AI Citations Never Say Your Name

September 14, 2026 · Vikram Jayanand

The Big Picture

Two numbers define the week. The People Also Ask box, long a reliable long-tail entry point for B2B content, is now almost entirely AI-generated, with fresh samples landing between 97% and 100%. And a new Semrush study found that 61.7% of AI citations never name the brand they came from, which means the standard GEO scorecard has been measuring the wrong thing. Underneath both stories sits the same uncomfortable point: exposure in AI surfaces is rising while attribution is falling, and Google's own admission that it made European search materially worse to satisfy the DMA is a reminder that the SERP you optimised for last quarter is not a fixed target.

Google AI Mode & AI Overviews

Interface changes worth watching

Google is testing thinner citation panels in AI Overviews, and both AI Overviews and AI Mode now surface pop-up link previews on hover. Separately, AI Mode is reported to run partially on Gemini 3.7 Flash, and Google confirmed dynamic expansion in late August, letting an AI Overview grow into a longer, AI Mode style answer where its systems judge a fuller response is more useful.

Key stat: PAA answers went from roughly 12% AI-generated to between 97% and 100% in fourteen months.

Audit which of your pages currently earn PAA visibility, then check whether those same URLs appear in the AI Overview citation panel for the same queries. If they do not, the traffic is already gone.
Thinner citation panels mean fewer visible slots. Treat citation-panel presence as a scarcer asset than it was six months ago, not a growing one.

Sources: Almost All People Also Ask Results Are Google AI Overviews, PAA Answers Are Now Almost All AI Overviews, September 2026 Google Webmaster Report

GEO / AEO / AIO

The ghost citation problem: cited is not the same as named

Semrush, working with Kevin Indig, analysed 3,981 domain appearances across four AI engines and found that 61.7% were ghost citations, meaning the AI used the page as a source link but never said the brand's name in the answer text. Only 13.2% of appearances were both cited and mentioned. Another 25.1% were brand mentions with no citation at all, where the AI named the brand but linked elsewhere.

The engine-level split matters more than the headline. When a brand appears in a Gemini answer, it is named in the text 83.7% of the time but cited as a source only 21.4% of the time. ChatGPT does close to the opposite: an 87% citation rate against a 20.7% mention rate. So a B2B brand optimising purely for ChatGPT citations may be building an asset that generates almost no name recognition, while a brand tracking only Gemini mentions may be reporting visibility that sends no referral traffic at all.

The study also found that short, conversational queries produce 30 to 50 times more brand mentions than long prompts, which trigger more citations but fewer mentions. That is a useful steer for anyone building a GEO prompt set: the prompts your buyers actually type are short, and short prompts are where naming happens.

Your personas are not seeing the same answers

Profound ran 71,147 responses across ChatGPT, Claude and Gemini between 11 and 24 August, varying only the persona described in the prompt. Responses to different personas for the same underlying question shared only about one in four brand mentions and one in five citations. Income, age, gender and occupation all shifted which brands got named and which sources got pulled.

For B2B, where the buying committee spans a practitioner, a director and a CFO asking materially different versions of the same question, this means a single AI visibility score is close to meaningless. You need the score per persona.

Semrush expands its AI Visibility Index

Semrush's expanded 2026 index now covers 126 million US AI search prompts from January through April, across ChatGPT, Gemini, Google AI Mode and AI Overviews. The concentration is stark: in News and Media the top three brands held 82.9% of total category visibility, and in Consumer Electronics 76.9%. Winner-take-most dynamics in AI answers appear stronger than in the classic ten blue links.

Key stat: 61.7% of AI citations are ghost citations, where the source is linked but the brand is never named in the answer.

Split your AI visibility reporting into two columns, citations and named mentions, and track them separately per engine. A single blended visibility number hides the failure mode.
Build your GEO prompt set from short, conversational buyer questions rather than long analyst-style prompts, since short queries are where brand naming actually happens.
Run the same prompt set through two or three distinct buyer personas and compare. Expect roughly 75% of the brand mentions to differ.

Sources: Why 62% of AI citations don't lead to brand mentions, Do Answer Engines customize responses to different personas?, Semrush Releases Expanded 2026 AI Visibility Index

Traditional SEO & Organic Search

Google says it made European search worse, on purpose

On 8 September Google rolled out its DMA compliance overhaul across European markets and described it as the largest reduction in quality of service in its 29-year search history. The changes follow a 460 million euro self-preferencing fine in July. The new layout places a third-party vertical search engine at the top of results flanked by two more, and strips real-time pricing from hotel and flight carousels. Comparison and vertical search services gain prominence over the businesses whose services they list.

The immediate impact lands on travel and local commerce, but the structural signal applies to any B2B brand with EU traffic: Google is willing to reshape the results page significantly and fast when compelled, and a layout-dependent organic strategy is fragile by design.

131 SEOs say relevance and links still run the table

Cyrus and Dawn Shepard published the 2026 Google Ranking Factors Expert Survey on 9 September, drawing 13,665 data points from 131 practitioners. Content relevance came top at 57.1%, backlinks second at 54.8%, and content quality third at 47.6%. Search intent match scored highest of any single factor. Meta description finished last. Links from trusted, topically relevant domains with real visitors scored highest within the backlink group, and spammy backlinks ranked among the strongest negative signals. Shepard's summary: reports of the death of backlinks have been exaggerated.

Mueller: programmatic SEO can cost you trust at the site level

Responding on Bluesky on 7 September, John Mueller said programmatic SEO often leads to a site that's either spam, borderline spam, or low quality, and that when it does, Google's systems have possibly lost faith in your site providing good value to users based on the old pages. The key word is site, not pages. Mueller added that recovery tends to take time and significant effort, and that the damage can persist well after the generated pages are deleted. Relevant for any B2B team that shipped a few thousand templated location, integration or comparison pages during the AI content rush.

Content marketing performance hits a 12-year low

Orbit Media's annual survey of 1,042 content marketers found only 14% say their blog delivers strong results, six points below the previous 12-year low and roughly half the 26% reported in 2022. AI adoption among respondents topped 92%. Orbit's reading is that the decline tracks the abandonment of the practices historically linked to strong performance, including original research, influencer collaboration, keyword research, paid promotion, formal human editing and consistent analytics use. No respondent reported using all eight.

One for the research pile

Google DeepMind, with UMass Amherst and UT Austin, published a paper proposing Autoregressive Ranking (ARR), where a single LLM generates a ranked list of document identifiers directly, replacing the dual-encoder plus cross-encoder pipeline most search systems use. It is a research paper, not a product announcement, and nothing in it says Google Search uses the approach. Also worth noting: Search Console's page indexing report is missing a chunk of June data, apparently a reporting bug rather than an indexing event.

Key stat: Only 14% of content marketers say their blog delivers strong results, the lowest in 12 years, while AI adoption among them passed 92%.

If you run EU traffic, re-screenshot your top commercial SERPs now and store them. The pre-DMA layout is no longer a valid baseline.
Backlinks remain a top-two expert-rated factor. If your 2026 plan swapped digital PR for GEO tooling, this survey is an argument for keeping both.
Pull a list of every templated page set you shipped in the last 18 months and assess it against Mueller's comment before Google does.

Sources: Google says DMA changes in EU resulted in worst degradation of search quality ever, Google's top ranking factors, according to 131 SEO professionals, 2026 Google Ranking Factors Expert Survey, Google Can Lose Faith In Sites Based On Low Value Programmatic SEO Pages, Content marketing success falls to 12-year low, Google DeepMind Develops New AI Search Ranking Model, Google Search Console Indexing Report Missing Old June Data

New Protocols & Standards

Google adopts IAB Tech Lab's ECAPI for the Data Manager API

The most concrete standards news of the week came from the ads side. Google is making its Data Manager API universal by adopting the IAB Tech Lab's ECAPI standard, and extending Data Manager itself beyond Google Ads into Google Analytics and Display & Video 360. For B2B teams with first-party data spread across a CRM, a MAP and a warehouse, a shared conversion API standard reduces the per-platform integration tax that has made server-side measurement expensive to maintain.

llms.txt: broad adoption, still thin proof

Adoption keeps climbing, with roughly 844,000 sites now publishing an llms.txt file, including Anthropic, Stripe, Vercel and Cloudflare. What has not changed is the evidence base. There is still no confirmation that the major AI search crawlers reliably fetch the file, and last week's reminder from John Mueller that the crawlers requesting markdown on his test sites were SEO tools rather than AI bots still stands. The practical consensus this week is to treat llms.txt as a site-level summary that complements page-level schema rather than replacing it, and to keep the investment proportionate to the evidence.

Schema markup remains the better-evidenced half of the pair. If you have to choose where developer time goes, structured data on comparison, pricing and product pages beats a hand-maintained llms.txt.
If you are rebuilding server-side conversion tracking this quarter, check whether ECAPI support changes your vendor shortlist before you sign.

Sources: Google expands Data Manager and Meridian measurement tools, Drive profitable growth with new data and measurement tools, llms.txt, Schema Markup, and Technical GEO: What Actually Works in 2026

Google Ads Updates

Exact and phrase match are being tested inside AI Mode

This is the week's most consequential paid change for B2B. Google is running what Ginny Marvin confirmed on 4 September as a small experiment letting standard Search campaigns built on exact and phrase match keywords serve text ads inside AI Mode, limited to cases with explicit and direct user intent. Until now the only routes into AI Mode placements were AI Max, Performance Max for Search, or broad match with Smart Bidding. For B2B advertisers who hold tight match-type control precisely because their qualified query set is small and expensive, this is the first path into AI Mode inventory that does not require surrendering targeting precision.

AI Max auto-upgrade is running now

The automatic migration of Search campaigns using campaign-level broad match or standalone Automatically Created Assets to AI Max runs through 30 September, rolling gradually and migrating in place with equivalent settings. Dynamic Search Ads are on a longer leash, with in-account notices starting this month, reminders in January 2027, and automatic migration in February 2027.

Amazon starts selling ChatGPT ads

On 10 September Amazon began offering ChatGPT Ads through Amazon DSP as a managed service, on both CPC and CPM, with Delta Vacations among the first US pilot brands. ChatGPT ads reportedly hit a $1 billion annualised revenue run rate in late August. Separately, OpenAI is reported to be blocking rival AI tools, including Adobe, from advertising competing AI products inside ChatGPT. Conversational ad inventory is consolidating fast, and it is being sold through the DSPs B2B teams already use.

Measurement gets a first-party data score

Google Ads is adding a Data Strength Uplift metric that quantifies the value generated by an advertiser's first-party data setup. Meridian, Google's open-source MMM platform, gains agentic capabilities for auditing data quality and guiding model building, and Meridian GeoX for causal geo-experiments is now generally available globally. Smaller items: a Video Campaign Group beta for video and Demand Gen campaigns with aggregated reach and frequency controls, and reporting segmentation by Loyalty program members.

Key stat: Exact and phrase match keywords can now serve text ads in AI Mode under a live Google test, a placement previously reachable only via AI Max, PMax for Search, or broad match.

If you have been holding out of AI Max to protect match-type control, watch this test closely. It may give you AI Mode reach without the migration.
Confirm your AI Max auto-migration has completed correctly before 30 September, checking negative keyword lists, brand exclusions and landing page mapping specifically.
Conversational ad inventory is now buyable through Amazon DSP. Worth a line in the FY27 plan even if you do not test it this quarter.

Sources: Google Ads Testing Serving Restrictive Match Types In AI Mode, Exact and phrase match keywords gain AI Mode ads in Google test, Google sets AI Max migration timeline for Search campaigns, Amazon pilots ChatGPT Ads through its DSP, Amazon Ads announces new integration with ChatGPT, OpenAI reportedly blocks rival AI tools from advertising in ChatGPT

Suggested Actions This Week

Split your AI visibility report into citations and named brand mentions, per engine. With 61.7% of citations never naming the brand, a blended score is telling you something you cannot act on.
Re-run your GEO prompt set through two or three distinct buyer personas from your ICP. Expect about three quarters of the brand mentions to change, and report the weakest persona, not the average.
Pull your current PAA visibility and check whether those same URLs appear in the AI Overview citation panels that replaced them. Where they do not, rewrite for citation extractability: clear definitional passages, named statistics, and dated sources.
Audit every templated page set you shipped in the last 18 months against Mueller's programmatic SEO comment. Site-level trust damage is much more expensive to fix than a few thin pages.
If you protect exact and phrase match for lead quality, monitor Google's AI Mode match-type test. It may be the first AI Mode entry point that does not require moving to AI Max.
If you run EU traffic, capture screenshots of your top commercial SERPs now. The post-DMA layout is the new baseline and the old one is gone.

← Back to Insights