81% of Marketers Still Call This "SEO". They're Not Wrong, and That's the Problem.

A survey of 343 US marketing decision-makers found that 81% describe their internal AI-search-visibility work as SEO, and only 27% have formally adopted any term beyond it.

The usual reaction to that number is one of two things. Either it is proof that the GEO and AEO vocabulary is vendor noise and the industry has sensibly ignored it, or it is proof that most marketers are asleep at the wheel while the ground moves under them.

Both readings are wrong, and in the same way. They treat the label as the thing being measured.

Here is the position the evidence actually supports: the 81% are right about their own work. The tactics, the craft, the quality bar — all of it still applies, and the vendors telling you to throw it out are selling something. And because they are right about the tactics, the naming hides a structural change that is real, is not a tactic, and is where teams are losing.

This is not a "SEO is dead" piece. The evidence published this summer does not support that claim, and several of the most-quoted datasets point the other way. It separates three things that keep getting bundled: what is continuous with SEO, what is genuinely new, and what the difference means for how a team is organized and measured.

---

Where the 81% comes from, and who produced it

Two attributions belong in the first paragraph rather than the last.

The survey was produced by Fractl, a content-marketing agency that sells AI-search services into the market it is measuring. The Search Engine Land piece carrying the numbers on August 24 was written by Fractl co-founder Kelsey Libert as a contributor column, with that relationship disclosed in line. Search Engine Land is the outlet here, not the researcher. The disclosure is the right practice; it does not make the instrument neutral.

The sample skews practitioner: 44% managers, 30% individual contributors, 14% directors, 10% C-suite, 2% VPs. Industry mix runs 27% B2C and e-commerce, 20% media and publishing, 19% B2B tech. Read "marketers" as "mostly managers and ICs at those kinds of companies", because that is what the sample is.

The aggregate also flattens a split that is more interesting than the aggregate. Among C-suite respondents, 28% use GEO and 17% use AEO — roughly three times the rate reported by individual contributors. Whatever is happening with terminology is happening at the top of the org chart first, and the people doing the work are still filing it under SEO.

That split is the shape of a specific organizational failure: the executive who reports the work has adopted a vocabulary implying a new discipline, while the team executing it has not changed what it does. When the reporting language and the operating model diverge, the gap fills with budget.

One more number from the same survey needs its base stated before anyone quotes it. Respondents reported that an average of 24% of their search or content budget now goes to AI visibility, with 82% allocating at least something. That is a share of the narrower search-and-content pot, not of total marketing spend, and the average is taken across a base that includes the roughly one in six respondents contributing a zero.

Source: Search Engine Land, AI search is here, but marketers still call it SEO (August 24, 2026)

---

The case that the 81% are right

Start here, because the continuity case is stronger than the industry's rhetoric suggests, and any argument that skips it is unserious.

The traffic base has not moved the way the headlines imply. Brainlabs gave Digiday GA4 data covering 54 of its own client accounts between January 2025 and April 2026. Organic search sessions across the set fell 10.5%, from 140.1 million to 125.4 million. Sessions referred from AI platforms rose 163%.

Now look at the bases rather than the percentages. That 163% is growth on roughly 200,000 AI-referred sessions a month against organic volume of under 15 million a month — about 1.3% the size. A 163% increase on 1.3% of traffic does not offset a 10.5% decline in the other 98.7%, and nothing in the dataset claims it does. The channel that pays the bills for most of these advertisers is still organic search, by two orders of magnitude.

Authority still mostly behaves like authority. A separate Fractl analysis, published in Search Engine Land on August 17, ran 96 industry prompts 15 times each against GPT-4o, Gemini 2.5 Flash and Claude Sonnet 4.6, producing 4,320 responses and more than 8,500 brand references mapped against Ahrefs authority metrics. It classified about 5% of brands as underexposed relative to their search authority and about 4% as overperformers, describing the remaining roughly 91% as behaving as expected.

Be precise about that 91%: it is not a correlation. It is the arithmetic remainder after subtracting two self-defined outlier buckets, published with no correlation coefficient, no outlier threshold, no collection date and no methodology appendix. Directionally interesting, not verifiable. But directionally it says what the 81% believe: the brands that earned search authority mostly show up.

A page a human built still beats a machine-generated one. Google Research's own generative UI paper, arXiv 2604.09577, ran generated interfaces against the single top Google Search result and won 90% of pairwise human comparisons on a 92-prompt set sampled from LMArena, dropping to 73.5% on Google's separate information-seeking set. That is the number that got quoted. The number that did not: in the same study, generated interfaces won only 35.3% of matchups against websites built by Upwork freelancers paid $100 to $130 for three to five hours of work — pages explicitly instructed to ignore SEO. On the information-seeking set, 30.0%.

A page somebody built in an afternoon for the price of a nice dinner beat Google's generative UI roughly two times in three. Those freelance builds are the floor of human-built web pages, not a proxy for a maintained tool page or a decade-old calculator. The generated interface lost to the floor.

And the format work still works. Kevin Indig's listicle study, published on his newsletter Growth Memo on August 31 and republished by Search Engine Land on September 2, found at least one listicle in Google's organic top 10 for 55.1% of 59,984 valid SERPs drawn from 60,000 US English desktop queries. In a separate crawl of 3,000 currently ranking listicles, pages with old or unreadable dates showed 56% lower odds of a top-3 placement.

Freshness signals, clear structure, a format that matches intent. That is an SEO checklist, and it is still predictive.

So when 81% of practitioners say "this is SEO", they are describing their actual work accurately. Nothing above requires a new discipline. Most of it requires the old one, done properly.

Sources: Digiday, In Graphic Detail: How AI search has impacted the web traffic of over 50 advertisers (August 25, 2026) · Search Engine Land, The AI visibility index: Which brands are vanishing from AI search? (August 17, 2026) · arXiv, Generative UI: LLMs are Effective UI Generators (February 24, 2026) · Search Engine Land, What makes a good listicle? (September 2, 2026)

---

What the label hides: three layers, and page work touches one

Here is the part the naming debate obscures.

In classical search there was effectively one system between your page and the user: an index and a ranking function. You optimized the page, the index picked it up, the ranking function placed it, the user saw a link. One pipeline, one feedback loop, and one reporting surface that told you how it went.

That is no longer the shape. Three distinct stages now sit between a page and an answer, and they fail independently:

Stage one is memory. What the model encoded during training, before it ever saw your page. Nothing you publish this quarter touches it directly.

Stage two is retrieval. Which queries the system fans out, and which documents it pulls. This is where page optimization operates — and only on the document side of it. You control the document. You do not control the fan-out.

Stage three is rendering. What the system builds out of what it pulled. Whether your document is named, quoted, summarized without attribution, or silently used to inform a sentence that credits nobody.

Page optimization, the thing every SEO team knows how to do, operates almost entirely at stage two. That is not a criticism of it. It is a statement about coverage. A team can execute stage two flawlessly and be invisible because of stage one, or be retrieved and then dropped at stage three, and no SEO reporting stack will distinguish those two failures from each other or from simply not ranking.

That is the structural change. It is not a new tactic. It is a change in the number of independent things that must go right, and in how few of them you can see.

---

The evidence at each layer is uneven, and the confidence runs the wrong way

The quality of evidence at each stage is inversely related to how confidently it is being sold.

At stage one, the most-cited finding comes from geoSurge, an AI-visibility vendor that sells what it calls corpus engineering. Its July 2026 research found that brands held in a model's top-10 category recall appeared in fan-out queries 55.7% of the time against 17.4% for brands not recalled — a 3.2x gap across 1,416 brand-level observations from 66 US prompts run 60 times each between May 29 and June 9, 2026.

Read the noun carefully. The study measured brand appearance in fan-out search queries. It did not measure whether those brands were named or cited in the answers a user reads. Memory was measured on one model and the queries were issued by Gemini 3.5 Flash, which the authors themselves flag, calling their number a conservative estimate. US prompts only. The authors describe the result as an association in exploratory data, not a proven cause.

And the extension from queries to answers runs against other available evidence. Separate measurements by the agency Victorious, by researchers at Georgetown and the University of Virginia, and by FrictionAI have each found that a model's familiarity with a brand does not predict that brand appearing in the answers users actually see.

A vendor selling corpus engineering published research concluding that corpus memory is what matters. That is not disqualifying. It is a reason to check what was measured.

At stage three, the evidence is Google's own, and it is the strongest of the three because it has something resembling a control group — the freelancer comparison above. It points away from the prevailing anxiety.

At stage two, where your existing team already operates, the advice is cheapest. Format for extraction. Date your pages. Keep the structure clean. The downside is zero.

The pattern across all three: the claims get more confident as the evidence gets thinner, and the confidence is highest exactly where somebody is selling the remedy.

There is one question that cuts through most vendor pitches in this category. Ask it before signing anything: does your number describe queries the system issued, or answers the user read? Much of the current literature quietly answers the first while charging for the second.

Sources: Search Engine Journal, The House Doesn't Publish Its Tells (August 21, 2026) · Search Engine Journal, Google's Generated Interfaces Could Compete With Tool Pages (August 21, 2026)

---

The failure mode is measurement, not tactics

If the tactics carry over and the layers are new, the place a team actually loses is the reporting layer. Three patterns from the last month show exactly how.

1. An agency's clients are not the web

Brainlabs' 54-client dataset showed organic down 10.5%. The same agency has separately reported that across the top 40,000 US websites, Google organic traffic fell about 2.5% in 2025 year over year, calling it a modest decline rather than a collapse. That is roughly a quarter of the drop measured inside its own client roster.

Both figures are Brainlabs'. Only one of them is about the web. The 54 accounts are a single agency's clients: self-selected, weighted toward large advertisers and retail, aggregated session-weighted so a handful of high-volume properties can move the headline on their own. No control group, no published methodology. And Brainlabs sells AI search optimization services, so a finding that organic is collapsing while AI referrals engage at a premium is directly flattering to that service line.

The same study also found the effect is heavily category-dependent. Brainlabs' own 1.35-million-keyword UK study in July 2026 put AI Overview coverage between 6.9% in energy and 43.9% in petcare. Within the newer client dataset, fitness, fintech, insurance and consumer packaged goods were hit hardest; retail, beauty and entertainment were least affected. There is no single number for the web in there.

2. The quotable engagement stat is not what it sounds like

The most-shared figure from that dataset is that referrals from ChatGPT, Copilot, Gemini and Perplexity triggered GA4 key events at 1.5 times the rate of organic search visits.

Key events in GA4 are configured per property by each advertiser. A key event can mean a purchase. It can also mean a newsletter signup, or scrolling to the bottom of a page. This is not a purchase-conversion comparison and should not be reported as one. Brainlabs also concedes the AI referral figures may run low because some AI-referred visits land in GA4 as direct traffic, which means the denominators behind the engagement comparison are not clean either.

3. When the reporting surface goes quiet, people fill it with inference

In August, Google shipped a spam update. The first-party record is short: the Search Status Dashboard lists an "August 2026 spam update" that started on August 18, 2026 and ran 2 days and 16 hours; Google said on LinkedIn it was "a normal spam update... for all languages and locations", and confirmed to Search Engine Roundtable that it was not a link spam update and not a site reputation abuse update. That is the entirety of it.

On August 25, Search Engine Journal published a report by Roger Montti collecting practitioner claims that the update had specifically hit sites built on fully automated AI publishing workflows. The named sources are two X accounts in the Japanese-language SEO community, anonymous posters on BlackHatWorld, and unnamed members of a private Facebook SEO group. No sample size, no measurement window, no dataset. The same article reports that some practitioners in that Facebook group said their AI-generated content improved after the update.

Google has not said this update targeted AI-generated content. The step from "normal spam update" to "this was the AI content update" is an inference made by practitioners, not a statement made by Google.

This is what a measurement vacuum produces. Not silence — a confident narrative assembled from selection-biased anecdote, because practitioners who lose rankings post about it and the unaffected mostly do not.

Sources: Search Engine Journal, Reports Indicate Google's Spam Update Focused On SEO AI Content (August 25, 2026) · Google Search Status Dashboard, August 2026 spam update (August 18, 2026) · Brainlabs, AI Overviews and declining keyword volume (July 20, 2026)

---

What is continuous, what is new

Stripped of vocabulary, here is the honest division.

Continuous with SEO — keep the team, keep the budget line, keep the craft:

• Crawlable, parseable, well-structured pages. Still the cheapest thing under your control.
• Freshness and dating. Indig's crawl found old or unreadable dates associated with 56% lower odds of a top-3 placement.
• Format matching intent. Listicles held a top-10 slot on 55.1% of the queries he sampled.
• Earned authority. Roughly nine in ten brands in Fractl's analysis behaved as their search authority predicted.
• Real pages with proprietary data, live state or a login. Google's own study says a human-built page still wins.

Genuinely new — and not solvable with page work:

• Cross-engine fragmentation. The buried figure in Fractl's August 17 analysis is that 77% of brands appeared in only one of the three models tested. Visibility in one assistant tells you very little about the others. Nothing in classical SEO has this shape; there was one index.
• The query-versus-answer gap. Being retrieved is not being cited. Three independent efforts have found that model familiarity with a brand does not predict its appearance in answers users read.
• Problem-framed queries. Fractl's own stated thesis, which cuts against its reassuring headline: brands with strong search authority still vanish from AI answers when a user describes a problem rather than naming a category. That is the gap where the work is not a rerun of SEO.
• No reporting surface you control. There is no equivalent of a rank-tracking baseline that is both neutral and complete. Every number in circulation was produced by somebody selling into the category.

The naming question is a labeling problem. Whether the work is different is a measurement problem. Only one of them has consequences.

---

What this means for how a team is organized

Three changes, in order of how much they cost.

Keep one team, not two. Nothing in the evidence justifies splitting an SEO function into an SEO function and a GEO function. The stage-two work is the same work, done by the same people, and a second team creates a second reporting line arguing for its own existence. The brief changes, not the headcount.

Add one measurement the old stack cannot produce: your own prompt set. Not a vendor's index. Your buyers' actual questions, phrased the way they phrase them, including the problem-framed versions where authority stops predicting presence, run repeatedly across more than one engine because the cross-engine overlap is low. This is the only instrument that distinguishes "we were not retrieved" from "we were retrieved and not named". Run it on a fixed cadence into storage you control, so a move tells you something. Tracking that consistently across engines and over time is the problem Genmark AI was built to solve, and it is also entirely possible to start by hand with a spreadsheet and fifty prompts.

Change what gets reported upward, not what it is called. If the C-suite has adopted GEO and AEO at three times the rate of the people doing the work, the correction is not a vocabulary rollout. It is putting two numbers side by side in the same deck: organic sessions, and presence in your own prompt set. The first is where the traffic still is. The second is the one no SEO report can produce. A team that reports both does not need to settle the naming argument.

---

The caveats, stated properly

Every figure in this piece is vendor-produced or platform-produced, with one partial exception. Fractl sells AI-search services and produced both the survey and the authority analysis. Brainlabs sells AI search optimization and produced the traffic decline. geoSurge sells corpus engineering and produced the memory finding. Google designed, evaluated and published the generative UI research. Indig is independent of the vendors he measures, but his SERP data came from SE Ranking, a commercial rank-tracking vendor whose product the scale of the dataset flatters, and the Search Engine Land republication does not disclose that.

The samples are small relative to the confidence of the percentages. 343 survey respondents, practitioner-skewed. 54 GA4 accounts, 48 of them US or UK. 1,416 brand-level observations from 66 prompts.

Several datasets are stale by publication. The Fractl survey was fielded on or before July 21 and written up August 24. Brainlabs' collection ended in April 2026. The geoSurge observations came from a twelve-day window at the end of May. Indig's SERP data was scraped in a single 47-hour window, and his own comparison of January 15 to August 24, 2026 found only 52.5% of queries kept the same position-one URL — on a surface that turns over half its top result in seven months, a two-day snapshot is a photograph, not a trend line.

What none of this establishes. It does not establish that AI search is displacing organic traffic at any particular rate for any particular business; the two Brainlabs figures differ by a factor of four and both are its own. It does not establish that model memory causes citation, which the geoSurge authors decline to claim. It does not establish what Google's August spam update did. And it does not establish that adopting or rejecting the term GEO changes any outcome — nobody has measured that, and the survey respondents themselves reported that they discount terminology fluency (9% named it as a credibility signal) and reward case studies with measurable results (34%).

---

The takeaway

The 81% are describing their work accurately. The tactics carry over, the craft still pays, and much of what is sold as a new discipline is the old one with a markup. Anyone telling you to discard the SEO function is arguing against the available evidence, including their own.

But the reason the label survives is the reason it is dangerous. It implies one pipeline with one feedback loop, and there are now three stages that fail independently, only one of which your team's instruments can see. The work did not change much. The number of blind spots did.

The fix is not a rename. It is one additional measurement, run on your own questions, on a cadence, in a place you control. Everything else in the debate is vocabulary.