In-depth coverage of AEO strategy, AI citation tracking, technical implementation, and real-world results from client work across 5 LLMs.
Most SEO changes are evaluated by shipping them and watching traffic. That method cannot distinguish a real effect from a seasonal shift, an algorithm update, or ordinary variance. Here is the testing methodology that can.
Google confirmed the crawl budget threshold at roughly one million unique pages changing weekly or ten thousand changing daily. Below that, crawl budget is not your constraint and optimising it is a waste. Above it, here is how to find where the waste actually is.
Every migration, domain change, and URL restructure produces a redirect map. Most are built manually in spreadsheets and most contain errors that cost rankings for months. Here is the process with the scripts that catch what a spreadsheet cannot.
Programmatic SEO generates pages from structured data and templates. When it works it captures long tail demand no hand written page could cover. When it fails it produces the exact scaled thin content Google now penalises explicitly.
Crawl delays run roughly 40 percent longer for JavaScript heavy sites, and Search Console data suggests around 80 percent of single page applications have crawl budget waste from pages queued for rendering that never get rendered. The framework you chose is making this decision for you.
WordPress is well suited to SEO out of the box and ships with several defaults that actively work against it. Most of the meaningful work is turning things off rather than installing more.
Images are usually uploaded by whoever wrote the post, named whatever the camera called them, and sized however the design tool exported them. That is why image optimisation remains one of the least contested opportunities available.
The phrase SEO copywriting still carries the smell of keyword stuffing and 500 word filler. The actual craft is narrower and harder: making the answer findable without making the prose worse.
Anyone giving you a single number is guessing or selling. The realistic answer depends on four variables, and knowing which one is binding on your site tells you more than any average timeline ever will.
Core updates now arrive roughly every three months rather than twice a year. Recovery runs three to six months and often requires the next update to register. Most sites spend that window fixing the wrong thing because they never established what they were hit for.
The free tools cover more than most people assume, and the expensive ones are worth it for a narrower set of jobs than their marketing suggests. Here is the honest split, including which metrics inside paid tools you should ignore.
Mobile first indexing means the mobile version of your page is the version that gets indexed and ranked. Content you hid on mobile for design reasons is content Google may not see at all.
Most SEO advice assumes a team, a budget, and a developer. If you are running a business and doing your own marketing, the useful question is not what works but what works first with a few hours a month.
The duplicate content penalty does not exist and never did. What does exist is signal dilution, wasted crawl budget, and Google choosing a different version of your page than the one you wanted ranking.
Winning a featured snippet means Google displays your answer directly on the results page. That is visibility and it is frequently a click you no longer get. The tactics still work and the reason for pursuing them has changed.
Google grades you on the 75th percentile of real Chrome users across a rolling 28 day window. A perfect Lighthouse score means nothing if a quarter of your visitors on mid range phones have a worse experience, which is exactly the situation on most sites.
The standard SaaS content playbook produces a blog ranking for hundreds of informational terms and a product nobody finds through search. The problem is not execution. It is that the playbook optimises for the wrong part of the funnel.
Most SEO reports lead with keyword rankings, which are the metric least connected to revenue and most easily flattered. Here is what to report instead, and how to handle the awkward conversation when traffic falls while performance improves.
Hreflang gets the attention because it is fiddly and produces visible errors. The decision that actually determines outcomes is the domain structure underneath it, and that one is expensive to reverse.
Search Console is the only source reporting what Google actually did with your site rather than what a tool estimates. Most people open the performance report, look at total clicks, and close it. The useful information is one filter away.
Well optimised category pages typically generate several times more organic revenue than individual product pages, because they rank for head terms and catch buyers earlier. Most stores treat them as a product grid with a heading.
Editorial.link surveyed 518 SEO professionals and found digital PR rated most effective by 48.6 percent, three times ahead of guest posting at 16 percent. Meanwhile roughly 96 percent of web pages have no external backlinks at all, which is the actual scale of the opportunity.
Most content strategies are publishing calendars with a keyword column added. A strategy decides what not to write, which is the part that determines whether the work compounds or accumulates.
Whitespark and BrightLocal's annual survey puts Google Business Profile signals at 32 percent of local pack ranking weight, the largest single category. A controlled study by Sterling Sky found GBP posts moved rankings by zero across 441 keywords. Most local SEO advice has these backwards.
Architecture is where large sites hide their cheapest wins. Reorganising internal links moves rankings without writing a single new word, and most sites have hundreds of pages sitting too deep to be crawled properly.
This is the first in a new comparison series. Rather than a head-to-head against a named competitor, this one addresses the comparison most teams actually face first: should we buy a tool, or is this simple enough to build ourselves with a spreadsheet and some API calls.
On-page SEO advice has accumulated twenty years of sediment. Some of it was never true, some was true in 2012, and a small number of elements do most of the work. Here is the separation.
This is the first entry in a new reference series. Rather than another explainer, this is a working glossary — short, direct definitions for the terms that come up constantly across AI search, AEO, and GEO discussions, each linked to a deeper guide where one exists on this blog.
Most keyword research produces a spreadsheet sorted by search volume, which is the least useful ordering available. Volume tells you how many people search. It says nothing about whether any of them want what you sell.
This is the first in a series of short posts correcting specific, common misconceptions circulating about AI search. Starting with the most persistent one: the idea that there is a discoverable, gameable ranking algorithm behind ChatGPT citations the way there is behind Google's organic search. There isn't, and treating it like there is leads to bad decisions.
Most audits produce fifty pages of findings and no prioritisation. This one runs in dependency order, because crawlability failures make every downstream fix pointless and almost every audit checks them last.
This is an illustrative, composite walkthrough built from the mechanics covered throughout this blog, not a single named client engagement. It maps the realistic sequence of changes a mid-size B2B site can expect over a 90-day AEO effort, and just as importantly, what does not move yet at each stage.
Traditional attribution captures around a quarter of the B2B buying journey. 6sense found 80 percent of deals go to the vendor already favoured before first contact. If your AEO reporting is a traffic chart, you are measuring the smallest and least decisive part of what is happening.
This is the first post in a new series organized by job title rather than industry. If you are the entire marketing department — writing the content, running the socials, occasionally touching the website code, and reporting straight to the founder — most AEO advice was not written with you in mind. This is the version that was.
Adobe measured AI referral traffic to US retail sites rising 693 percent year over year during the 2025 holiday season, with AI referrals converting 31 percent better than other traffic. For most consumer brands, the pages winning those queries belong to retailers rather than the brand itself.
Running the same prompt against ChatGPT, Perplexity, Claude, Gemini, and Google AI Overviews every week, at scale, across thousands of tracked queries, sounds simple until you actually try to build it. Here is the real architecture behind it, including the parts that took three rebuilds to get right.
Knowing content decays is not the same as having a system that catches it. Most refresh programmes fail because they are organised around publication date, which is the least useful signal available.
Server log analysis across a meaningful sample of sites found GPTBot referencing sitemaps on roughly three out of four initial domain visits, compared to just over half for Googlebot. PerplexityBot showed an even stronger dependency, crawling sitemap-listed URLs several times more often than equivalent pages it found only through links. Most sitemaps are still built entirely around Google.
Professional services sell expertise that AI systems now dispense for free in summary form. That sounds like an existential problem and is closer to a positioning one, because the queries that lead to an engagement are not the queries AI answers well.
A page with zero organic clicks and ten ChatGPT citations is not a pruning candidate — but most content audits, run purely from Google Analytics, would flag it for deletion anyway. The freshness guide in this series covers keeping content current. This is the other half: knowing when a page is actively hurting you and should be killed, not refreshed.
Forrester found 94 percent of B2B buyers used AI during their most recent purchase, with 47 percent building internal business cases inside AI tools before contacting a vendor. Your best research is behind a form, which means it contributes nothing to the case they are building.
Google's own list of schema types still worth investing in after its 2026 rich-result cleanup names Event alongside JobPosting as one of the categories that maps to genuine, ongoing business use rather than SERP-gaming. Most companies running webinars and conferences have never implemented it.
Review queries trigger an AI Overview 86.3 percent of the time according to Seer Interactive. Most companies treat responses as reputation management theatre. They are actually publicly indexed text where you get to state your side of a specific factual claim.
ChatGPT alone handles roughly 2.5 billion prompts a day, and a meaningful share of those touch professional decisions — including which recruiter or job to trust. Most agencies and employer career pages are answering none of it, because the content that wins AI citation in hiring looks nothing like a job board listing.
Directories and marketplaces have thousands of pages they did not write, describing products they do not control, updated by third parties. Almost every AEO principle assumes you own your content. Here is what changes when you do not.
On May 6, 2026, Google announced five new features for AI Mode and AI Overviews explicitly aimed at expanding outbound linking — including a new "explore new angles" section that surfaces distinct, in-depth articles beyond the synthesized answer itself. It does not resolve the fundamental tension of a system that answers the query directly, but it is a genuine, concrete structural change worth understanding.
AirOps analysed more than 245,000 prompts that brands were actively monitoring and found they peak at six to seven words. Real AI search queries routinely run past ten. Teams are measuring a part of the curve where the traffic no longer lives.
Seer Interactive measured X versus Y queries triggering an AI Overview 95.4 percent of the time, the highest rate of any query class they tested. If you sell software, your comparison pages are your highest leverage AEO asset and they are probably written as sales collateral.
Landing pages are usually noindexed, thin by design, and built entirely around a form. Meanwhile Forrester found 94 percent of B2B buyers now use AI during their purchase. The pages carrying your sharpest messaging are excluded from the channel where buyers are shortlisting.
Google added Google-CloudVertexBot to its official list of crawlers in 2026 — a new user-agent specifically for sites owners that build Vertex AI Agents and want that agent crawling their own site on their behalf. It is one of a growing list of newer, purpose-specific AI crawlers most robots.txt files have simply never been updated to address at all.
How much does it cost is one of the highest intent questions anyone can ask, and it is increasingly asked to an AI rather than typed into a search box. If your pricing lives behind a contact form, something else is answering on your behalf.
A case study saying a client saw significant improvements in efficiency contains zero extractable information. The format is one of the strongest citation opportunities a company has and most of them are written in a way that guarantees nothing can be pulled from them.
Buyers evaluating software ask compatibility questions before almost anything else, because an integration gap is a hard disqualifier. Most companies answer this with a wall of logos, which is the least extractable format available.
Most About pages are a founding story and a stock photo of a team. The page is doing a structural job nothing else on your site can do, and treating it as brand storytelling wastes the only place your organisation is formally defined.
On March 13, 2026, Perplexity launched Premium Sources, giving its Computer and Finance products direct, built-in access to paywalled data from CB Insights, PitchBook, and Statista. For any brand competing for visibility on market-sizing, funding, or industry-benchmark queries, this changes exactly who you are competing against for the citation.
A comment section full of real questions and real answers is some of the most citable content you can host. A comment section full of spam and one word replies is thin content stapled to every page you publish. Most sites have the second and think they have nothing.
Definition content is the most directly citable format there is, because a question starting with what is has exactly one shape of answer. It is also the format with the most competition and the lowest tolerance for padding.
You spent a week producing a twenty minute technical walkthrough. To a text based retrieval system, that page contains a title, a thumbnail, and an embed code. Everything you actually said is not there.
Nobody writes a changelog for visibility. It exists because users need to know what changed. That accidental honesty is exactly what makes it citable, and most companies bury it behind a login or ship it as a modal that no crawler will ever see.
Google's own documentation states plainly that AI Overviews only show when its systems determine the response is genuinely additive to classic search, and that they "often don't trigger" as a result. This is a distinct question from citation optimization. Before you can optimize for being cited inside an Overview, it helps enormously to understand which of your queries produce one at all.
If your blog has four hundred posts behind an infinite scroll archive, most AI crawlers can see the first ten. The rest may as well not exist. This is one of the few AEO problems where the content is genuinely fine and the delivery mechanism is doing all the damage.
A byline is a string of text. An entity is something a machine can connect to other things it already knows. Most sites publish the first and assume they built the second.
Putting your blog on blog.example.com instead of example.com/blog is a decision usually made by whoever set up the CMS. It determines whether your content builds one authority signal or two weaker ones, and it is expensive to reverse.
A page pulled out of context is just text. Breadcrumb markup is the cheapest way to tell a machine what section that text came from and what it relates to, and it takes about an hour to implement across a whole site.
At Google I/O 2026, Google announced that OpenAI, Kakao, and ElevenLabs are adopting SynthID, its content-provenance watermarking system, alongside expanded content-transparency and verification tools across Search, Gemini, Chrome, Pixel, and Cloud. This is arriving at the same moment global regulation is formalizing AI-content disclosure requirements. Together, they change what an E-E-A-T-credible authorship signal actually needs to demonstrate.
Your best research is probably a PDF. Your whitepaper, your annual report, your technical spec. AI engines can read some of them and are effectively blind to others, and the difference has almost nothing to do with how good the content is.
Migration checklists were written for Googlebot. They cover redirects and canonical tags and sitemap resubmission. None of them tell you that different AI engines will notice your migration at wildly different speeds, and that the gap between them is where citations quietly disappear.
Comparison tables are the highest value content type you can publish for AI search, because a direct comparison is exactly what a lot of buyer queries are asking for. They are also the content type most often built in a way machines cannot read.
Status codes are the only direct conversation your server has with a crawler. Most sites are sending the wrong ones without knowing it, and a soft 404 is worse for AI citation than a hard one.
At Google I/O 2026, Google announced Universal Cart, a shopping layer spanning Search, Gemini, YouTube, and Gmail, alongside updates to the Universal Commerce Protocol and the Agent Payments Protocol. The strategic bet is explicit: shoppers will increasingly describe full requirements in natural language, and brands will influence AI shopping agents through structured product data, not through on-site merchandising.
At Google I/O 2026 on May 19, Google announced that Search can now build a custom interface on the fly — interactive visuals, tables, graphs, simulations, even persistent "mini apps" for ongoing tasks — assembled in real time using Gemini 3.5 Flash and Google Antigravity. This is a genuinely new frontier for AEO, and the standard question "was my page cited" does not fully apply to it anymore.
On May 27, 2026, Google extended its Preferred Sources feature — previously confined to Top Stories — into AI Overviews and AI Mode. When a reader has personally marked your site as preferred and you get cited in a generated answer, your link now carries a visible badge, and Google has confirmed preferred status is becoming an input into citation frequency itself, not just a display label.
On May 15, 2026, Google published its first official, consolidated guide to optimizing for AI Overviews and AI Mode. It tells site owners to stop worrying about llms.txt, content chunking, and AI-specific rewriting. It does not say AEO or GEO work is pointless. Here is exactly what the guide says, section by section, and what it means for a real content programme.
On June 5, 2026, Google updated its Search Central documentation to formally name AEO and GEO as legitimate SEO services, and, in the same set of pages, published a public framework for auditing whether a vendor selling those services is credible. It is a genuinely useful accountability tool, whether you are hiring an agency or evaluating your own internal AEO programme.
As of May 7, 2026, FAQ rich results no longer appear anywhere in Google Search, for any site, in any query type or language. This followed the earlier 2023 restriction that had already limited the feature to a narrow set of government and health properties. Yes, this changes the calculation. No, it does not mean you should delete your FAQPage schema.
Seventy-two countries now have some form of AI policy, according to OECD tracking. All fifty US states have introduced AI legislation. South Korea and Japan enacted dedicated AI laws in 2025. This is not a single global trend converging on one standard. It is a genuinely fragmented landscape, and a content team operating across three or more major markets is already, in practice, operating under three or more different sets of expectations.
Agent-readiness audits now score sites from 0 to 100 across five categories and sixteen checks — Discoverability, Content, Bot Access Control, API/Auth/MCP/Skill Discovery, and Commerce. Most sites score under 30. This is the complete tutorial: what each check actually tests, why it matters, and exactly how to fix it, with working code for every single item.
The EU's Article 50 marking rules take effect August 2, 2026. India's IT Rules amendment, defining and regulating "synthetically generated information," has been in force since February 20, 2026, with a strict three-hour takedown window. If your content team publishes AI-assisted material across multiple markets, the labeling rules are no longer the same everywhere, and treating them as one global standard is a mistake.
ChatGPT Search converts each 128-token chunk into a numerical vector and scores it against the query vector using cosine similarity, completing the entire scoring pass across every candidate chunk in 100 to 200 milliseconds. That single comparison, run at GPU speed across thousands of chunks, is the moment your content either enters the answer or gets discarded. Understanding the math changes how you write.
Microsoft Bing confirmed officially in December 2025 that large language models group near-duplicate URLs and select a single representative page to cite. If your page is not chosen as that representative version, Bing's own documentation states plainly that it is unlikely to be cited or summarised in AI-generated answers at all. A canonical tag is your only lever over which version wins, and it is a weaker lever than most teams assume.
Vercel analysed over 500 million GPTBot fetches and found zero evidence of JavaScript execution. None. GPTBot, ClaudeBot, and PerplexityBot download your raw HTML, extract whatever text is already there, and move on immediately. If your content only exists after a client-side JavaScript bundle finishes running, every one of these crawlers sees an empty shell, while Google, using headless Chrome, sees your site perfectly.
A 30-day study across twelve production sites found GPTBot revisits high-traffic pages roughly every 2.4 days, ClaudeBot every 6.8 days, and Google-Extended every 14 days on a near-metronomic schedule. Google Analytics shows none of this activity, because AI crawlers do not execute the JavaScript that GA4 depends on to register a visit. The only place this behaviour is visible at all is in your raw server access logs.
News Corp reportedly earns around $50 million per year across its portfolio for licensing content to AI companies. OpenAI has struck at least 18 disclosed publisher deals. But the entire licensing market, by structural necessity, pays only the brand-name corpus with genuine negotiating leverage, and every serious 2026 industry analysis concludes the long tail of small and mid-size publishers will see no meaningful revenue from this at all.
Three separate EU AI Act enforcement mechanisms activate simultaneously on August 2, 2026. This is not primarily a concern for AI model builders. If your team uses generative AI tools to produce marketing content, tracks AI citations, or operates in the EU market at all, several of these obligations reach directly into ordinary content and marketing operations.
Ahrefs studied 75,000 brands and found that web mentions correlate with AI Overview visibility at 0.664. Backlinks correlate at just 0.218. A mention with no hyperlink attached carries roughly three times the predictive weight of a traditional do-follow link. This single finding rewrites two decades of digital PR strategy.
SparkToro's clickstream data shows 68.01% of US Google searches now end without a single click, up from 60.45% just two years ago. Yet brands cited inside an AI-generated answer earn 35% more organic clicks than competitors who are not cited. SEO is not dying. The version built entirely on collecting clicks from ranked links is the part that is dying.
Ahrefs studied 863,000 keyword SERPs and four million AI Overview URLs in February 2026 and found that only 38% of cited pages now rank in the traditional top 10, down from 76% in mid-2025. Pages without traditional ranking authority can still win citations if they are structured cleanly enough for AI extraction. Schema markup is the cleanest structuring tool available, and this is the single guide that ties every type together.
On May 13, 2026, Google quietly shipped a native AI Assistant channel into GA4. Before that date, the overwhelming majority of properties were systematically under-counting AI-originated visits by 30 to 40%, lumping ChatGPT and Perplexity sessions into Direct or generic Referral traffic. The native fix helps. It is not the whole solution, and AI referral traffic converts 4.4 times higher than standard organic, which makes getting this measurement right a budget-allocation problem, not a reporting nicety.
Google's May 2026 generative AI guide states directly that unique, compelling, useful content will influence AI search visibility more than any other single recommendation in the document. It calls the opposite of this "commodity content" — and in a world where any competitor can generate a competent generic article in seconds, that distinction has become the entire game.
AI Overviews now appear on roughly 25% of all Google searches. Organic click-through rate drops 61% when one appears above your result. But pages cited inside the Overview earn 35% more organic clicks and 91% more paid clicks than competitors left out. Getting cited is no longer optional. Here is exactly what determines whether you make it in.
GEO is not a marketing buzzword someone invented to sell a course. It is a peer-reviewed academic discipline, formalised in a 2024 paper by researchers at Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi, who tested nine content strategies across 10,000 queries and measured the exact visibility lift each one produced. The headline finding: targeted optimization can improve AI visibility by up to 40%.
ChatGPT handles over 2.5 billion messages a day. There is no position one. There is no SERP. Either your brand is in the generated answer, or it is invisible to that conversation entirely. ChatGPT SEO is the discipline of earning that inclusion, and it runs on different rules from the SEO playbook most marketers already know.
A founder asked a caller where they heard about the company. "ChatGPT," the caller said. The founder had never optimised for it on purpose. That story repeats across thousands of small businesses right now. You do not need an agency or a six-figure budget to start showing up. You need about a week of focused work and a clear understanding of what ChatGPT actually rewards.
Google's own May 2026 documentation confirms, in plain language, that AI Overviews and AI Mode issue multiple related searches across subtopics and data sources before generating a single response. This is why a page that ranks fifth for a broad query, but contains one exceptionally specific paragraph, can still earn a citation the broad-query ranking alone would never predict.
Your content does not compete at the level of pages. It competes at the level of 256-token chunks. A RAG system retrieves roughly 20 to 50 candidate passages, scores each against the query using vector similarity, reranks the top ten with a cross-encoder, and feeds the top three to five to the generator. Every AEO decision you make either helps or hurts your content at one of those stages.
The finding that 44.2% of AI citations come from the first 30% of a page is not a quirk of editorial preference. It is a structural consequence of how transformer attention accumulates across layers. The earliest tokens in a sequence accumulate computational advantage that grows with model depth. Every BLUF writing rule has a mechanistic explanation in the architecture.
79% of ChatGPT responses draw from training weights, not live web search. Perplexity draws from live web retrieval on every query. Claude defaults to training data and uses search as a deliberate tool call. The AEO implication is direct: the same content change produces different citation outcomes on different engines, because each engine is a different kind of memory system.
When your product is called "NotionCue Prompt Tracker" on your website, "Prompt Tracking Tool" on G2, and "the tracker" in your blog posts, AI engines are processing three different entity candidates. They may merge them correctly or may not. Inconsistent naming is not a branding problem. It is an entity resolution failure that reduces citation confidence across every engine simultaneously.
As of May 7, 2026, FAQ rich results no longer appear anywhere in Google Search, for any site, in any query type or language. This followed the earlier 2023 restriction that had already limited the feature to a narrow set of government and health properties. Yes, this changes the calculation. No, it does not mean you should delete your FAQPage schema.
Half of all AI-cited content is under 13 weeks old. A page with zero quality change earns 3.2x more citations at day 14 than at day 91, purely because AI retrieval systems structurally favour recent content over older content. This is not a quality problem. It is a maintenance problem, and it has a systematic fix.
AI engines are the new comparison sites. A buyer asking "best CRM for a 10-person sales team" gets a synthesised shortlist directly in the AI answer, bypassing the affiliate roundup entirely. The affiliate and comparison sites still earning AI citations in 2026 are the ones AI engines trust as a source, not the ones they replace.
Most AEO optimisation is based on best-practice implementation, not measured experiments. You add FAQPage schema and hope citations improve. You rewrite your opening paragraph and watch for weeks. Without a structured before-and-after test, you cannot separate improvement from coincidence, or know which specific change moved the number. Here is the AEO testing framework that generates actionable data.
AI citation patterns for a new product are established within the first 30 to 60 days of launch. The sources AI engines cite first, the announcement page, the press coverage, the community discussion, become the baseline that subsequent retrievals anchor to. Brands that engineer their launch content for AI citation before day one compound that advantage for months.
On June 5, 2026, Google updated its Search Central documentation to formally name AEO and GEO as legitimate SEO services, and, in the same set of pages, published a public framework for auditing whether a vendor selling those services is credible. It is a genuinely useful accountability tool, whether you are hiring an agency or evaluating your own internal AEO programme.
AI engines cannot see images. They read alt text, captions, surrounding paragraphs, and ImageObject schema. A product page with no alt text and no ImageObject schema is invisible in visual AI search, regardless of image quality. Here is the complete implementation that makes images citable.
Help center articles answer the exact questions buyers ask AI engines before making a purchase decision. "Does this tool support Salesforce?" "What is the file size limit?" "How do I migrate from Competitor X?" These are live queries running through ChatGPT and Perplexity right now. Your knowledge base either earns those citations, or your competitor does.
When a learner asks ChatGPT "best online course for Python machine learning," AI engines pull from schema.org Course data, authoritative course review platforms, and educational institution signals. EdTech brands without Course schema and instructor E-E-A-T credentials are invisible in these answers. Here is the implementation that earns citations.
Most AEO prompt sets are built wrong. Teams track branded queries ("What is [product]?") and miss the pre-purchase moments where AI engines are making buying decisions without the buyer ever typing a brand name. The prompts that matter are the ones buyers run before they know you exist. Here is how to build the right prompt set.
On May 15, 2026, Google published its first official, consolidated guide to optimizing for AI Overviews and AI Mode. It tells site owners to stop worrying about llms.txt, content chunking, and AI-specific rewriting. It does not say AEO or GEO work is pointless. Here is exactly what the guide says, section by section, and what it means for a real content programme.
AI engines apply their most cautious citation patterns to financial content. A consumer-electronics brand earns citations through review density alone. A fintech brand needs regulatory credentials, named authors with verifiable qualifications, rate disclosures with effective dates, and methodology pages, before a citation is possible. Here is what that means for your content programme.
AI engines bypass language barriers. Perplexity and Gemini pull facts from English-language sources and synthesise answers in French, German, or Portuguese in real time. Sites that only translate without localising lose 431% in AI citation eligibility against properly localised competitors. Hreflang is necessary but nowhere near sufficient.
A 10% increase in G2 reviews correlates with a 2% increase in AI citation rates, social proof signals are being processed directly by citation algorithms. For comparison queries like "best running shoes under £100," AggregateRating schema determines whether your product makes the AI's shortlist. Here is the exact implementation with zero common errors.
AI engines cannot listen to your podcast. They read transcripts, metadata, and schema. A show with 50,000 downloads per episode but no structured transcript earns zero AI citations. A show with 400 listeners and corrected HTML transcripts plus AudioObject schema competes directly for citation in ChatGPT, Perplexity, and Google AI Overviews. The fix installs in one week.
Seventy-two countries now have some form of AI policy, according to OECD tracking. All fifty US states have introduced AI legislation. South Korea and Japan enacted dedicated AI laws in 2025. This is not a single global trend converging on one standard. It is a genuinely fragmented landscape, and a content team operating across three or more major markets is already, in practice, operating under three or more different sets of expectations.
YouTube is now cited 200 times more than any other video platform in AI-generated answers. Subscriber count and view count have near-zero correlation with AI citation rate. A 400-view explainer video with a clean transcript and VideoObject schema competes on exactly the same terms as a two-million-view video that has neither.
Reddit accounts for roughly 40% of AI citations across ChatGPT, Gemini, and Claude, the single most-cited source in every major engine. Not your website. Not your blog. An anonymous forum. Here is what that means for your brand and what you can actually do about it.
Traditional SEO requires six to twelve months of authority building before a new domain earns competitive rankings. AEO requires two to three hours of technical setup before you start appearing in AI answers. The overlap between AI citation sources and Google's top-ten results is only 12%. You do not need to beat established competitors at SEO to beat them at AEO.
Adding inline citations to named sources improves AI visibility by up to 40%, per the Princeton GEO research paper. But being cited is even more valuable than citing others. Original data, benchmark studies, proprietary product analytics, customer survey findings, earns citations from AI engines that synthesised content from the same sources simply cannot compete with.
Three separate EU AI Act enforcement mechanisms activate simultaneously on August 2, 2026. This is not primarily a concern for AI model builders. If your team uses generative AI tools to produce marketing content, tracks AI citations, or operates in the EU market at all, several of these obligations reach directly into ordinary content and marketing operations.
The EU's Article 50 marking rules take effect August 2, 2026. India's IT Rules amendment, defining and regulating "synthetically generated information," has been in force since February 20, 2026, with a strict three-hour takedown window. If your content team publishes AI-assisted material across multiple markets, the labeling rules are no longer the same everywhere, and treating them as one global standard is a mistake.
Copilot converts at 17x the rate of direct traffic for subscription conversions, the highest conversion rate of any AI platform measured. It runs entirely on Bing's index, not Google's. If you have never opened Bing Webmaster Tools, Copilot cannot cite you. Here is the six-step fix.
Claude is more conservative than ChatGPT about which sources it cites, and more likely to simply decline when evidence feels thin. It uses Brave Search for web retrieval, not Bing or Google. It values original analysis over synthesis. Here is what that means in practice.
News publishers expect search traffic to drop 43% within three years, per Reuters Institute. AI Overviews are already cutting first-result CTR by 58%. The publishers maintaining visibility are not fighting AI, they are engineering their content to be the source AI cites. Here is how.
When a developer asks ChatGPT how to integrate with your API, ChatGPT retrieves from documentation, not from your marketing site. When a power user asks Perplexity about your product limits, Perplexity cites your docs. Product documentation is not a support cost, it is an AI citation surface that most SaaS teams have not optimised.
Google's May 2026 generative AI guide states directly that unique, compelling, useful content will influence AI search visibility more than any other single recommendation in the document. It calls the opposite of this "commodity content" — and in a world where any competitor can generate a competent generic article in seconds, that distinction has become the entire game.
Triple-schema stacking, FAQPage plus Article plus HowTo on one page, produces 1.8x more AI citations than Article schema alone. HowTo schema makes each step independently extractable. Here is the complete implementation with copy-paste JSON-LD.
Bidirectional internal linking increases AI citation probability by 2.7 times versus one-way links from pillar to spoke. The anchor text you use on internal links is a topical signal, not a housekeeping task. Here is the complete internal linking framework for AI search.
A broken schema block is ignored entirely by AI crawlers, not partially processed. A page with malformed FAQPage schema gets zero schema benefit, the same as a page with no schema. These are the ten errors that appear most often in AEO audits, with exact fixes.
Most content teams add AEO to their backlog and treat it like a channel to launch. It does not work that way. AI citation authority is built through weekly iteration, not campaigns. Here is the workflow that fits into an existing content calendar without replacing it.
Harvard Business Review documented two concurrent revolutions in early 2026: the move from SEO to GEO, and AI agents beginning to act as buyers. Voice commerce is heading for $80 billion globally this year. The brands that appear when an AI agent evaluates a purchase will win customers the brand never directly interacted with.
Google's own May 2026 documentation confirms, in plain language, that AI Overviews and AI Mode issue multiple related searches across subtopics and data sources before generating a single response. This is why a page that ranks fifth for a broad query, but contains one exceptionally specific paragraph, can still earn a citation the broad-query ranking alone would never predict.
Voice queries now account for 27% of all searches. There are more voice assistants active globally than there are people on Earth, 8.4 billion devices. Voice returns exactly one answer. If it is not you, it is a competitor. Here is how to become the spoken answer.
77.67% of legal queries trigger a Google AI Overview, the highest rate of any industry. Healthcare queries trigger AI Overviews 88% of the time. But the citation bar in YMYL sectors is significantly higher. Generic content gets blocked. Authoritative, credential-anchored content earns citations at a premium.
The AEO tool market matured fast. Enterprise platforms like Profound, Conductor, and BrightEdge offer broad coverage but start at thousands per month. Lighter tools fill specific gaps. Here is how to match tool to use case without overspending on features you will not use.
Speakable schema is the only structured data type that explicitly tells Google which page sections to extract for text-to-speech. Most sites implement FAQPage and Article schema but skip Speakable entirely. Here is how to implement it correctly and why it matters beyond just voice search.
Most agencies are being asked about AEO before they have a repeatable process for it. The brands winning the AI citation conversation with clients are not the ones with the most complex workflows, they are the ones with the clearest measurement framework and a consistent execution sequence.
AI Overviews appear on 57% of long-tail, high-intent shopping queries. Over 60% of consumers now begin product research with an AI assistant. Traditional SEO got you to the category page. AEO gets you into the answer when a buyer asks which product to buy.
Gemini powers both AI Overviews and AI Mode, the two surfaces that together appear on over 55% of all Google searches. Getting cited by Gemini is different from getting cited by ChatGPT or Perplexity. Here is exactly what Gemini prioritises and what you need to change.
95% of the sub-queries that trigger AI citations have zero traditional search volume. Standard keyword research tools do not find them. Here is a different research method built specifically for AI search, and the prompt types that earn the most citations.
60% of searches now end without a click. AI Mode has a 93% zero-click rate. Traffic is not disappearing, it is being absorbed inside AI answers. The brands that thrive in a zero-click environment are not fighting it. They are building for it.
Enterprise teams that invested in AEO early saw 2 to 3x improvement in AI share of voice within 12 months, per McKinsey 2026 data. HubSpot's own AEO programme produced a 1,850% increase in qualified leads. The ROI is real. The problem is proving it with data your current analytics setup cannot easily produce.
60% of US and EU searches now end without a click. ChatGPT processes 2.5 billion daily prompts. If you are still measuring search success only through rankings and organic traffic, you are measuring the wrong channel. This is what AEO is, how it works, and where to start.
Three terms. One goal. The industry cannot agree on what to call the practice of getting cited in AI-generated answers. Here is exactly what each term means, where they overlap, and which lens actually helps you build a better programme.
Most AEO problems are not content problems. They are access problems, structure problems, or entity signal problems. This 27-point audit finds them in priority order, crawl issues first, content issues second, entity issues third.
44.2% of all AI citations come from the first 30% of a piece of content. The single highest-impact structural change in AEO is also the simplest: put the answer first. Here is exactly how to do it across every content type you publish.
The average brand appears in fewer than 1 in 5 AI answers on relevant queries. AI visibility score is the metric that tells you exactly where you sit, and the four levers that move it.
Perplexity visits about 10 pages per query but only cites 3 to 4. It runs a six-stage RAG pipeline that most brands are not building for. This is what actually happens between your URL and a Perplexity citation.
ChatGPT retrieves pages for 85% of queries but only cites 15% of those pages. 32.9% of all cited pages were found through fan-out sub-queries with zero traditional search volume. Here is what that means for your content strategy.
51% of B2B software buyers now begin their research with an AI chatbot more often than with Google. That number was 29% in April 2025. If your brand is absent from those AI answers, you are invisible to the majority of your potential pipeline before they ever reach your website.
Google Analytics still does not show which AI engine sent a visitor. GSC does not separate AI Mode from organic clicks. Most teams are flying blind on a channel that converts at 4x the rate of standard search. Here is how to actually measure AEO performance.
AI Mode replaced the traditional results page entirely in May 2026. Only 14% of URLs cited in AI Mode overlap with AI Overview citations. Neither maps to traditional organic rankings. Here is exactly how the retrieval works and what you need to change.
AI engines do not retrieve content from domains. They retrieve content from entities they trust. Your brand is an entity in an AI knowledge graph, whether you have built it deliberately or not. Here is what that graph contains about you and how to fix it.
A site with thirty tightly interconnected articles on one subject gets cited far more often than a site with one comprehensive guide. This is how AI engines measure topical authority, and this is the architecture that builds it.
96% of AI Overview citations come from sources with strong E-E-A-T signals. The correlation between traditional domain authority and AI citation has collapsed to 0.18. Trust is now the primary eligibility filter for AI citation, and it works differently from how most teams build it.
Before your content reaches ChatGPT, Perplexity, or Claude as a cited answer, a bot has to fetch your page. Most sites are blocking the wrong bots without knowing it.
Traditional content gap analysis tells you which keywords your competitors rank for. AEO content gap analysis tells you which questions AI engines are answering with your competitors when you are the right source. The two processes are not the same.
When buyers ask ChatGPT, Perplexity, or Claude about your category, how often does your brand appear compared to competitors? AI Share of Voice is the metric that answers that question, and the average brand mention rate across AI engines sits at just 17.2%.
Three major studies, 300,000+ domains analysed, and one finding: llms.txt does not directly move AI citation rates. Here is what it is actually good for.
Google Analytics shows nothing wrong. The page still ranks. But your brand has vanished from AI-generated answers. This is citation decay, and it is the AEO problem most teams find out about six months too late.
Most teams count brand mentions, call it a metric, and wonder why nothing improves. Prompt tracking is not keyword rank tracking with a different interface. Here is how to build a set that produces signal.
Pages with well-implemented structured data earn AI citations at 2.8x the rate of unstructured equivalents. Here are the five schema types that drive that gap, in priority order, with copy-paste JSON-LD for each.
A content strategy living entirely on your own domain has a structural ceiling. Reddit, review platforms, Wikipedia, and editorial coverage are where AI engines do most of their sourcing.
The framing that AEO replaces SEO is wrong. The framing that AEO is just SEO with a different name is equally wrong. Here is exactly where they share foundations and where they have genuinely diverged.
Most content that fails to get cited is not bad content. It is content written for a human reader that an AI retrieval system cannot efficiently parse. The fix is structural, not qualitative.
Hallucination rates on brand-specific information run between 20% and 76% depending on the engine. For brands that have not built a systematic evidence layer, the AI's best guess is often wrong in ways that are commercially damaging and completely invisible unless you run the prompts yourself.
59% of ChatGPT searches involve local intent. AI Overviews appear in 92–97% of local informational queries. Most local businesses have no strategy for appearing in those answers. Here is what actually works.
We analysed 50,000 pages cited by ChatGPT, Gemini, and Perplexity. Here's what schema types appear most frequently.
A step-by-step breakdown of the llms.txt fix, BLUF rewrites, and schema additions that took ChatGPT citations from near-zero to 33K monthly sessions.
Which bots respect it, which ignore it, what format actually works, and why getting this wrong can silently exclude you from AI training pipelines.
OpenAI's April 2026 update changed how GPT-4o selects sources for commercial intent queries. Here's exactly what shifted.
Managing multi-language, multi-region AEO is a different challenge from single-domain work. Our guide to hreflang, regional llms.txt, and language-specific BLUF.
Perplexity and real-time AI crawlers actively penalise slow pages. Our data shows domains with Good CWV scores get cited 2.1x more often.