Google published a blunt answer to the answer engine optimization (AEO) and generative engine optimization (GEO) boom in May 2026. From its perspective, optimizing for generative AI search is still SEO. The company told site owners that its AI features still depend on core search ranking and quality systems. It also told them to ignore supposed shortcuts such as special AI markup, tiny “AI-ready” content chunks and pages rewritten for machines.
Three months earlier, Bing had launched an AI Performance dashboard that tracks how often a site is cited in generated answers. Bing called it an early step toward GEO tooling.
Both positions hold.
Google still needs a crawlable, indexable, useful page. Bing now reports when that page is displayed as a cited source in an AI answer. The work overlaps. The measurement does not.
A webpage now competes for four separate outcomes:
Retrieval: A search system discovers, retrieves and ranks it.
Attribution: An answer displays it as a cited source, and the citation actually supports the associated claim.
Selection: The response names or recommends the business in the right context.
Execution: A person or delegated agent can complete the next authorized action.
A company can win the first contest and lose the next three.
The old scoreboard stopped explaining the outcome
Rankings were never the whole of search, but they gave operators a useful chain of evidence. A page appeared for a query. A person clicked it. Analytics recorded the visit. The site converted some share of those visits.
Generated answers interrupt that chain.
Pew Research Center studied 68,879 Google searches conducted by 900 U.S. adults in March 2025. On pages classified as having an AI summary, users clicked a traditional result in 8% of visits. On pages without one, they clicked a result in 15% of visits. Only 1% of visits to pages with an AI summary produced a click on a cited source. Longer, question-shaped searches were more likely to produce AI summaries. Read Pew’s method and findings.
The researchers observed browsing behavior in March, then used a third-party scraping service in April to reconstruct the first results page for the same searches. They did not capture each participant’s exact live or personalized results page. This is an observed association, not proof that an AI summary caused the difference. It still establishes the operating problem: a rank position cannot tell you whether a generated answer appeared, displayed your page as a source, named your brand, selected another option or ended the search.
A page can keep its rank while failing to earn the searcher’s click.
SEO, AEO and GEO are working labels, not separate sciences
The industry does not use AEO and GEO consistently. Google folds both into SEO. Practitioners often use them interchangeably. The original GEO research defined a more specific problem: improving a source’s visibility inside a generated response, where citations appear with different prominence instead of occupying a numbered rank. Its authors tested a 10,000-query benchmark and separately ran a point-in-time experiment on Perplexity.ai. That was a useful early result, not validation across every answer engine operating today. Read the KDD 2024 GEO paper.
The labels become useful when each one names an observable failure:
SEO
Primary lens: Can a system discover, index, retrieve and rank the page?
Evidence to inspect: Indexing, impressions, rank, clicks and crawl health.
Common failure: The page is inaccessible, ambiguous or too generic to retrieve.
AEO
Primary lens: Does the response answer correctly, and do its citations support the key claims?
Evidence to inspect: Answer inclusion, factual accuracy, displayed citations and citation fidelity.
Common failure: The answer is wrong, unsupported or sourced elsewhere.
GEO
Primary lens: Is the source attributed or the business selected and framed in the right context?
Evidence to inspect: Cited URLs, citation fidelity, mentions, framing and selection share across a fixed prompt set.
Common failure: Better-corroborated sources construct the answer without you.
This is a working vocabulary, not a standards document. It should help a team diagnose where visibility failed. If it creates three departments selling three copies of the same content calendar, it has failed.
One buying question now produces four verdicts
Consider a buyer asking for the best payroll platform for a 20-person U.S. agency that also pays international contractors.
A vendor can rank well for “payroll platform” and never appear in the generated answer. That is an attribution or selection failure, not proof of an indexing problem.
The vendor can be cited for a tax definition while another company gets named on the shortlist. That is source visibility without commercial selection.
The vendor can make the shortlist and still lose because its supported countries, pricing, implementation time or contractor limits are hidden behind vague copy and a contact form. A browser agent cannot compare details that the site never states. A human buyer often will not wait either.
The same query has now produced a retrieval verdict, an attribution verdict, a selection verdict and an execution verdict.
Blending them into one “AI visibility score” destroys the diagnosis. A team cannot repair what the score refuses to distinguish.
Google is right about the foundation
Google’s current guidance is unusually direct. Generative features use retrieval-augmented generation grounded in Google’s index. They also use query fan-out, generating related searches to gather material before composing an answer. A page still has to be indexed and eligible for a normal search snippet before it can appear in Google’s generative features. Read Google’s official generative AI search guide.
That keeps the old foundations in force:
Crawlable pages
Clear site architecture
Original, non-commodity information
First-hand experience
Accurate titles and headings
Consistent entities, products and claims
Useful images, video and structured data where they fit
There is no secret “AI content” format that replaces those things.
Google explicitly says it ignores llms.txt for Search, including its generative features. A non-Google system may consume the file only if that system documents support. A team can also maintain one as a public-resource map. It is not access control, a training opt-out or a Google ranking tactic. Treating the file as a visibility spell is cargo cult SEO with a newer filename.
Google also warns against splitting content into tiny chunks for AI, generating a page for every fan-out variation or seeking synthetic mentions. Clear writing helps. Manipulative volume does not become legitimate because the target is an answer box.
The scoreboard really did change
Google’s foundation argument does not erase the need for new measurement. Its own Search Console now has generative AI performance reports for a subset of sites. Those reports separate impressions in AI Overviews, AI Mode and generative Discover experiences. Read Google’s Search Console announcement.
Bing goes further. Its AI Performance dashboard reports total citations, cited pages, grounding queries and citation trends across Microsoft AI experiences. Grounding queries are a sample of overall citation activity. Bing carefully notes that a citation count does not reveal rank, authority, placement or the page’s role inside an individual answer. Read Bing’s dashboard announcement.
That distinction matters. In Bing’s dashboard, a citation is evidence that the page was displayed as a source. It does not prove that the page supported a particular claim, that the system recommended the company or that the answer drove a sale.
The new scoreboard needs separate columns:
Retrieval: Did the page appear and rank for the query?
Attribution: Which pages were displayed as sources, was the answer accurate and did each citation support its associated claim?
Selection: Was the company named, framed or recommended in the right context?
Execution: Could the buyer or a buyer-authorized agent complete the next meaningful step?
No single percentage can carry all four meanings.
The practical operating loop
Run the work in this order.
1. Separate search access from training access
OpenAI documents separate controls for OAI-SearchBot and GPTBot. A site can allow the search crawler so its pages remain eligible for ChatGPT search while disallowing the training crawler. The settings express different choices. Read OpenAI’s crawler documentation.
Do not let one blanket robots rule make both decisions accidentally. Check the live file, CDN rules and bot protection before rewriting content.
2. Publish something a model could not invent without you
Commodity summaries provide less distinctive evidence to retrieve or attribute. Publish the test, the dataset, the implementation detail, the dated observation, the named method or the first-hand failure. Then state what it proves and what it does not.
Original information can give a search system a reason to retrieve the page and an answer system a reason to attribute it.
3. Make each important claim inspectable
Use normal headings and paragraphs. Answer the section’s question early. Attach the source, date, author and boundary close to the claim. Keep structured data consistent with what a reader can see.
This is not “writing for robots.” It is removing the ambiguity that frustrates readers, search engines and answer systems at the same time.
4. Repair the record beyond your own domain
Generated answers synthesize multiple sources. Your homepage cannot by itself correct a stale directory, an outdated review, inconsistent product documentation and five forum threads describing the same confusion.
Map the sources that appear for the fixed buyer questions. Correct records you control. Earn accurate coverage and references where you do not. Leave sources you cannot influence labeled as terrain, not unfinished work.
5. Make the next action possible
Google now tells site owners to prepare for agents that act on behalf of people by comparing specifications, inspecting the DOM, reading the accessibility tree and performing tasks such as booking. That makes semantic HTML, visible product details, current availability, clear pricing units and labeled controls part of agent readiness, not only conversion design.
The agent does not need a futuristic protocol to discover that a button has no label or a price has no unit. The failure is already visible in the page.
Do not buy three copies of the same strategy
SEO remains the retrieval foundation. AEO is a useful lens for answer accuracy and attribution. GEO spans attribution and selection inside generated responses. Agent readiness tests execution.
Run them as one evidence system with separate scoreboards.
Start with ten real buyer questions. Test them across ChatGPT, Claude, Gemini and Perplexity at a fixed cadence for one month, with repeated runs. Record the product and model when shown, browsing or search mode, date, location or account state, ranked pages, generated answers, displayed citations, citation fidelity, company selection and whether the next action worked. Preserve the raw captures. Generated answers vary across engines, time and executions, so one run is an anecdote, not a baseline. Read the 2026 ACL comparison. Read the words, not only the score.
Then fix the first broken handoff in the chain.
Your search presence now has four jobs: it must be retrievable, attributable, selectable and executable. A new acronym does not perform any of them for you.
FAQ
Is GEO another name for SEO?
For Google Search, the official answer is yes: its generative features use the same core ranking and quality systems, and Google treats AEO and GEO work for those features as SEO. Operationally, GEO is still a useful label for two distinct measurement problems: whether a source is attributed and whether a business is selected and framed in the right context.
What is the difference between AEO and GEO?
There is no settled industry standard, and the terms often overlap. A useful working distinction is that AEO examines answer accuracy and attribution, while GEO examines attribution and selection inside a generated response. The underlying repair work overlaps heavily.
Do I need an llms.txt file or special AI schema?
Not for Google Search. Google says it ignores llms.txt and requires no special AI markup. Maintain an llms.txt only for systems that document support or as a clear map of public resources. It is not access control or a training opt-out. Use supported structured data when it accurately matches visible page content, not as a guarantee of AI inclusion.
What should I measure first?
Measure a fixed set of real buyer questions with repeated runs across the engines your customers use. Record retrieval, attribution, selection and execution separately. Keep the raw, dated output and test conditions so a later change can be compared with evidence instead of memory.
Can anyone guarantee an AI citation or recommendation?
No outside practitioner controls the retrieval and generation systems. A credible engagement can guarantee the audit trail, the fixes it ships and the cadence of repeated measurement. It cannot guarantee what an independent engine will say next.
Jason Colapietro is the founder of Suede Labs AI. Suede runs one SEO, AEO and GEO practice because the work shares one source of truth, while each outcome keeps its own scoreboard. Read the public method or contact Jason at seo.suedeai.ai.