
Sign up to save your podcasts
Or


Show Notes
Microsoft Just Proved a Point About Search Today
Announcing Microsoft Web IQ — Bing Search Blog
https://blogs.bing.com/search/June-2026/Announcing-Microsoft-Web-IQ
Introducing AI Performance in Bing Webmaster Tools (public preview) — Bing Webmaster Blog
https://blogs.bing.com/webmaster/February-2026/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview
New AI Visibility Insights in Bing Webmaster Tools: Intents, Topics, Citation Share, and Compare — Bing Search Blog
https://blogs.bing.com/search/June-2026/New-AI-Visibility-Insights-in-Bing-Webmaster-Tools-Intents-Topics-Citation-Share-Compare
Jordi Ribas on Web IQ and how AI agents search — Search Engine Land
https://searchengineland.com/microsoft-releases-web-iq-powered-by-bing-but-designed-for-how-ai-agents-search-479194
Microsoft Web IQ gives AI agents Bing grounding APIs — Search Engine Journal
https://www.searchenginejournal.com/microsoft-web-iq-gives-ai-agents-bing-grounding-apis/577736/
The evolving role of the index: from ranking pages to supporting answers — Bing Search Blog
https://blogs.bing.com/search/May-2026/Evolving-role-of-the-index-From-ranking-pages-to-supporting-answers
Rank and AI Citation Aren’t the Same Number — Duane Forrester Decodes
https://duaneforresterdecodes.substack.com/p/rank-and-ai-citation-arent-the-same
The Machine Layer (book) — Amazon
https://www.amazon.com/Machine-Layer-Visible-Trusted-Search/dp/B0G2WZKM59/ref=sr_1_1
Show Notes: 81.8% of My “AI Assistant” Traffic Was Fake
Episode summary
Over two weeks, a brand-new website with zero promotion behind it logged thirty-three visits from AI assistants. Only six were real. The rest were lying about who they were, and the Googlebot numbers were worse. In this episode I walk through exactly what I found in my own server logs, how I proved each finding past the point of doubt, and the simple method you can run on your own logs this week to see your real numbers. We cover why a bot’s name is a claim and not an identity, the difference between bots that fetch you for a live answer and bots that crawl you to train tomorrow’s models, the one crawler I had to chase four different ways to nail down, and the one major player you structurally cannot measure at all.
What you’ll learn:
• Why the bot names in your analytics are a “claims to be” number, not a real one, and the one check that fixes it.
• The 81.8 percent spoof rate hiding in live AI-assistant traffic, and how the fakes gave themselves away.
• Why Googlebot showed 87 percent impersonation, and why that is an old story, not a new one.
• The difference between retrieval crawlers (today’s visibility) and training crawlers (whether the model knows you tomorrow).
• A repeatable, four-step way to settle any bot you cannot verify on the first pass.
• Why Gemini is the one source you cannot measure by name, and how that rhymes with Google’s old “(not provided)” move.
The numbers, at a glance:
• Live AI-assistant fetches: 33 claimed, 6 verified, 27 spoofed. An 81.8 percent spoof rate among the requests that could be checked.
• Googlebot: 799 claimed, 107 verified, 692 spoofed. Roughly 87 percent not Google.
• Most active verified crawlers: Anthropic’s ClaudeBot 166, Googlebot 107, OpenAI’s GPTBot 46, OpenAI’s search crawler 40.
• CCBot (Common Crawl): 20 claimed, 0 verified. Confirmed as impostors across four independent checks.
A reminder these are two weeks on one small, new site. The method is the point, not my totals.
The published IP-range lists (verify your own logs)
These are the first-party files each operator publishes. A request is only legitimate if its source IP falls inside the matching list. Each link goes straight to the source.
OpenAI
ChatGPT-User (live user fetch): https://openai.com/chatgpt-user.json
OAI-SearchBot (search / retrieval): https://openai.com/searchbot.json
GPTBot (training): https://openai.com/gptbot.json
Anthropic (one file covers all of their bots, including ClaudeBot and Claude-User)
https://claude.com/crawling/bots.json
Perplexity
Perplexity-User (live user fetch): https://www.perplexity.com/perplexity-user.json
PerplexityBot (crawler): https://www.perplexity.com/perplexitybot.json
Google (note: Google moved these to the /crawling/ipranges/ path in 2026, and the old URLs fail quietly)
Common crawlers, including Googlebot: https://developers.google.com/static/crawling/ipranges/common-crawlers.json
Special-case crawlers: https://developers.google.com/static/crawling/ipranges/special-crawlers.json
User-triggered agents: https://developers.google.com/static/crawling/ipranges/user-triggered-agents.json
Common Crawl
CCBot: https://index.commoncrawl.org/ccbot.json
Verification and proof resources
• Google’s full crawler and fetcher reference, where Google states that Google-Extended has no separate user agent and is a robots.txt control token, not a fetcher, and where Google itself warns the user agent string can be spoofed: https://developers.google.com/crawling/docs/crawlers-fetchers/google-common-crawlers
• Google’s guide to verifying that a request really came from Google, including the reverse-DNS method: https://developers.google.com/crawling/docs/crawlers-fetchers/verify-google-requests
• Common Crawl’s public index. Drop in a domain and a recent crawl to check whether your site is actually in the corpus. Use a wildcard, for example yoursite.com/*, so you are not just matching the homepage: https://index.commoncrawl.org/
Run it yourself: the four-step chase
When a bot will not verify on the first pass, do not stop at “unknown.” Do this:
1. Check the published IP list. Is the source address inside the operator’s ranges?
2. Check reverse DNS. Does the IP resolve back to the operator’s own hostname?
3. Check the corpus or index where one exists, like Common Crawl’s, to see if you were actually captured.
4. Run a WHOIS lookup on the raw IP to see who really owns it. Commodity hosting in random countries is your answer.
Four angles that agree is proof. One that does not is a thread worth pulling.
Try it and tell me.
Run this on your own logs and send me two numbers: your demand spoof rate, and your Googlebot one. I suspect the real story is in the spread between them.
More on the question of what happens to your content after the fetch: https://www.citationiq.com
Follow the show so the next episode finds you.
Resources and references
Query length, prompts vs. searches:
SimilarWeb data on how much longer AI prompts run than Google queries: https://officechai.com/ai/chatgpt-queries-17x-longer-than-google-searches-6x-longer-than-googles-ai-mode-similarweb-data/
Clickstream analysis on the gap between the typed prompt and the search the model actually fires: https://martech.org/chatgpt-growing-as-a-traffic-referrer-reshaping-search-behavior-report-says/
Study on prompt decomposition, multiple retrieval searches per prompt: https://searchengineland.com/chatgpt-search-prompts-data-463407
Longtail and ranking:
Why longtail is about specificity and search volume, not word count: https://www.yotpo.com/blog/long-tail-keywords-guide/
On long, specific phrases being easier to rank for at modest authority: https://www.w3era.com/blog/seo/long-tail-keyword-strategy/
On reading search volume as a starting point, not a verdict: https://www.outrank.so/blog/how-to-find-low-competition-keywords
Citation vs. organic overlap:
Moz, on most AI Mode citations not appearing in the organic results for the same query: https://thenextweb.com/news/ai-changing-seo-tools
ZipTie, on how few cited URLs land in Google's top ten: https://ziptie.dev/blog/how-different-ai-platforms-cite-the-same-source-differently/
Semrush AI Mode study, including heavy Perplexity-Google overlap: https://www.semrush.com/blog/ai-mode-comparison-study/
How input shape moves what gets surfaced:
Comparative analysis, AI sourcing shifting with the character of the query: https://arxiv.org/abs/2601.16858
Study on outputs shifting when prompts are rephrased: https://arxiv.org/abs/2509.08919
Book:
The Machine Layer: https://www.amazon.com/Machine-Layer-Visible-Trusted-Search/dp/B0G2WZKM59/ref=sr_1_1
Referenced in this episode:
When the Training Data Cutoff Becomes a Ranking Factor (Duane Forrester Decodes) https://duaneforresterdecodes.substack.com/p/when-the-training-data-cutoff-becomes The companion piece this episode builds on, where I first laid out the parametric-versus-retrieval distinction and what it means for timing.
How Perplexity finds and chooses its sources (Search Engine Journal) https://www.searchenginejournal.com/perplexity-ai-interview-explains-how-ai-search-works/565395/ Background on why Perplexity runs a live search on essentially every query rather than answering from memory.
Google's AI optimization guidance, and why AI Search is still Search (DemandSphere) https://www.demandsphere.com/blog/google-ai-optimization-guide-ai-search-is-still-search/ Support for the point that AI Overviews and AI Mode are served off the core Search index, not from Gemini's parametric memory.
Claude web search tool documentation (Anthropic) https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool Primary source showing Claude's web search runs as a tool the model invokes only when it decides a question needs it.
Manage public web access in Microsoft 365 Copilot (Microsoft Learn) https://learn.microsoft.com/en-us/microsoft-365/copilot/manage-public-web-access The admin control behind the point that, on Copilot, whether retrieval happens at all can be a tenant policy setting.
Stop Treating AI Visibility as One Problem (Duane Forrester Decodes) https://duaneforresterdecodes.substack.com/p/stop-treating-ai-visibility-as-one The earlier governed-visibility piece this episode zooms into, treating retrieval as one of three layers to manage.
ChatGPT search behavior, clickstream insights (Semrush) https://www.semrush.com/blog/chatgpt-search-insights/ The study behind the stat that ChatGPT's share of search-triggering sessions swung between roughly 15 and 66 percent as models updated.
Lost in the Middle: How Language Models Use Long Contexts (arXiv) https://arxiv.org/abs/2307.03172 The foundational research on models using long context unevenly, behind the point that being retrieved isn't the same as being used well.
How up to date is ChatGPT, and how knowledge cutoffs work (JustDone) https://justdone.com/blog/ai/how-up-to-date-is-chatgpt Context for the training-cadence point that providers now ship frequent point releases, each carrying its own cutoff.
The Machine Layer (Amazon) https://www.amazon.com/Machine-Layer-Visible-Trusted-Search/dp/B0G2WZKM59/ref=sr_1_1 My book, for the longer argument on why visibility, trust, and machine-readability are converging into one problem.
References from this episode:
When I mentioned Gerard Salton's SMART system at Cornell, the foundational vector space model work from the 1960s, here's the background on that: https://en.wikipedia.org/wiki/SMART_Information_Retrieval_System
The Netflix study I referenced on cosine similarity producing arbitrary results in embedding models, that's the 2024 paper from Steck, Ekanadham, and Kallus: https://research.netflix.com/publication/is-cosine-similarity-of-embeddings-really-about-similarity
The MTEB benchmark leaderboard, where you can see the performance spread across current embedding models: https://huggingface.co/spaces/mteb/leaderboard
Goodhart's Law, the "when a measure becomes a target it ceases to be a good measure" concept: https://en.wikipedia.org/wiki/Goodhart%27s_law
The vector index hygiene piece I referenced from last year: https://duaneforresterdecodes.substack.com/p/vector-index-hygiene-a-new-layer
The written version of this episode is available as the full article on this same Substack.
The Machine Layer - available at Amazon now.
The tools are deployed. The licenses are paid. And if you’re a senior SEO or GEO practitioner right now, you’re probably using AI every day - for drafts, for summaries, for first passes at content that used to take twice as long. That’s real productivity, and it’s not nothing.
It’s also not the return the investment is capable of producing. And the gap between what you’re getting and what’s available isn’t a tool problem. It’s a mode problem.
A peer-reviewed study published at the 2025 ASIS&T Annual Meeting by Tim Gorichanaz at Drexel University gives that problem a name (h/t to Shari Thurow for pointing me at this paper!). Analyzing 205 real-world ChatGPT use cases, Gorichanaz identified six distinct modes in which people actually use AI: Writing, Deciding, Identifying, Ideating, Talking, and Critiquing. The data came from Reddit and skews Anglophone, which limits its generalizability, but the taxonomy it produced maps uncomfortably well onto how most practitioners are actually working. Two modes dominate. Four are being left on the table. The four being left are the ones that determine whether AI makes you more strategically valuable or just faster at execution-layer work.
That distinction matters more right now than it has at any prior point in this industry’s history.
The Two Modes Everyone Defaults To
Writing was the largest category in Gorichanaz’s data at 47% of observed use cases - drafting, editing, summarizing, translating, generating. McKinsey’s 2025 State of AI survey confirms this at the enterprise level: the most commonly reported AI use cases are content drafting and information capture, and 63% of organizations using generative AI apply it primarily to create text.
Identifying - explaining something, answering a factual question, summarizing a document - was another 10% of the study’s data, and represents the other pillar most practitioners have built their AI workflow around. Research a topic, get a synthesis, move to the next task.
Together these two modes account for the overwhelming majority of how AI is being used, by practitioners and enterprises alike. Both have real value, yet neither is where the leverage is. And if your AI practice begins and ends there, you’re using an increasingly sophisticated tool to do work that was already being automated - just faster and at higher volume.
The other four modes (Deciding at 21% of Gorichanaz’s sample, Ideating at 9%, Talking at 8%, and Critiquing at 6%) are where the work becomes irreplaceable. They’re also where almost no practitioner has built a deliberate workflow, because nobody handed them one, and the pressure to show immediate output has consistently crowded out the space to develop one.
The Decisions You’re Still Making Alone
In the practitioner’s week, Deciding-mode questions are everywhere: which queries actually have AI visibility exposure worth prioritizing right now, whether a brand’s retrieval problem is a content architecture problem or a sourcing and signal problem, how to allocate effort across a portfolio when both SEO and GEO need attention and the budget doesn’t stretch to cover both fully, when to escalate a visibility concern to leadership versus when to fix it in the work before anyone asks.
Most senior practitioners are currently solving these questions with experience and intuition. That’s not a failure as experience and intuition are genuinely valuable, and no AI replaces them. But AI used deliberately in Deciding mode adds something experience can’t provide on its own: a structured pressure-test of the assumptions underneath the decision, applied before the decision hardens.
That requires more than a good question. Deciding mode requires giving the AI the relevant context (competitive landscape, current visibility posture, historical performance, strategic constraints) and then treating what comes back as a genuine input to the decision rather than a draft to be skimmed and set aside. It requires a workflow that doesn’t yet exist in most practitioners’ practice, not because anyone blocked it, but because no one built the time or structure for it either.
The same McKinsey data makes clear what that gap costs at scale: 88% of organizations use AI, but only 6% qualify as high performers generating meaningful enterprise-wide impact, and high performers are 3.6 times more likely to have fundamentally reworked their workflows rather than simply deployed tools into existing ones. The pattern holds at the practitioner level. Faster output from an unreconstructed workflow is not the same thing as better decisions from a restructured one.
The Gaps Nobody Briefed
For SEO and GEO practitioners, Ideating mode has a specific application that most are not using and most should be: mapping the entity and authority gaps the brand hasn’t recognized yet.
What angles of topical authority has the brand failed to establish that AI retrieval systems are currently filling from other sources? What community signals (forum discussions, aggregated reviews, third-party commentary) are shaping how LLMs represent the brand in response to category queries, and what would it take to shift them? What framings of the brand exist in model training data that the brand’s own content has never addressed or countered?
These are genuinely Ideating-mode questions. They’re also questions most practitioners have some version of in the back of their mind without a structured method for surfacing the answers. AI used in Ideating mode, not “give me five content ideas” but a genuine iterative exploration with deliberate constraints and real willingness to follow the output somewhere the team hasn’t already been, is one of the most direct methods available for finding those gaps before a competitor or a client audit finds them first.
The barrier isn’t capability. It’s the difference between a Writing prompt with a list output and an actual Ideating session. The first takes two minutes. The second takes twenty, requires a different posture toward the tool, and produces something that can’t be replicated by anyone who didn’t do it. That asymmetry is where practitioner value gets built in the current environment, and most practitioners are not claiming it.
The Honest Read Your Team Won’t Give You
This is the mode with the most direct application to daily practice and the most organizational resistance, because it requires using AI to find problems in work the practitioner or their team has already invested in.
Used properly, Critiquing is how a senior practitioner catches what internal review missed. The weak entity claim in a content strategy that sounds authoritative but isn’t backed by the sourcing AI retrieval systems actually trust. The gap between what the brand says about itself across owned properties and what a well-prompted LLM surfaces when asked a category question the brand should own. The assumed premise in a GEO recommendation that made sense six months ago and is now contradicted by how retrieval patterns have shifted.
That last application is not abstract. Running your own brand (or a client’s brand) through a structured AI Critiquing session before the next strategy cycle is exactly the kind of proactive work that separates practitioners operating at the judgment layer from practitioners operating at the production layer. It’s also the kind of work that changes the conversation with a client or a leadership team, because you’re surfacing problems before they become visible in the data rather than explaining them after the fact.
The reason Critiquing is underused isn’t a governance problem. It’s a disposition problem. Organizations and practitioners have broadly trained themselves to use AI to produce output, not to interrogate it. Reversing that habit is a choice, and it’s one of the more consequential choices available to a senior practitioner right now.
Rehearsal
The Talking mode in Gorichanaz’s taxonomy covers AI as a conversation partner, and for practitioners, the most valuable version of that is rehearsal f…
For roughly two decades, the SEO discipline operated on a quiet assumption that turned out to be one of its most valuable features. Guidance from one search engine traveled. If Google said sitemaps mattered, Bing said sitemaps mattered. If Bing said structured data deserved real effort, Google said the same. Practitioners optimized for Google with reasonable confidence that the work would carry across the other engines, and most of the time it did. That portability was not luck. It was the product of a structurally large overlap layer that the major search engines had jointly built, brick by brick, over twenty years.
That world doesn’t exist in LLM-land. The major providers train on different corpora, run different crawlers under different policies, route different queries through different retrieval systems, and apply different alignment processes that shape the final response in ways the upstream signals can’t predict. Guidance from any one provider, including Google’s guidance about its own Gemini products, is one data point. Practitioners carrying the SEO habit forward, the habit of treating one engine’s guidance as roughly the whole map, will optimize confidently for one platform and miss the others.
Sidebar: As I was finalizing this piece, Google published fresh guidance on optimizing for their generative AI features. Their framing is explicit: from Google Search's perspective, optimizing for AI search is still SEO. That framing is accurate for Google Search. It does not extend to ChatGPT, Claude, Perplexity, or any other LLM, and that is precisely the trap this article is about.
The shared standards that made SEO guidance portable
The era of portable guidance was built on actual collaboration, not coincidence. The Sitemaps protocol became the joint property of Google, Yahoo, and Microsoft in November 2006, when the three engines formally agreed to support a common protocol at version 0.90, building on Google’s earlier Sitemaps 0.84 from June 2005. Five years later, on June 2, 2011, the same three engines launched Schema.org, with Yandex joining shortly after, to create a common vocabulary for structured data markup. That was the announcement that got made on stage at SMX Advanced. I was on the Bing team at the time, and what struck me then is what still matters now. The engines were competitors, but they had decided that a shared vocabulary served them all. Webmasters got one set of rules. The web got cleaner data. The engines got better signals. Everybody won.
The pattern repeated with robots.txt, the 1994 convention that became RFC 9309 at the IETF in 2022, formalizing what every serious crawler already honored. And it repeated again, more recently, with IndexNow, the protocol Microsoft Bing and Yandex launched in October 2021. IndexNow is now supported by Bing, Yandex, Naver, Seznam, and Yep. Google has tested the protocol since 2021 but has not adopted it.
That overlap layer is exactly why Google’s guidance felt safe to follow even if you cared about Bing traffic. The signals the engines used were not identical, but the inputs they accepted, the protocols they honored, and the standards they advertised were. Optimization had a shared substrate.
Where the LLM stacks actually diverge
The LLM environment doesn’t have a shared substrate of comparable size. The differences are not cosmetic and they are not temporary. They are baked into how the systems are built.
Start with training data. OpenAI has signed disclosed licensing deals with News Corp worth up to $250 million over five years, Axel Springer at roughly $13 million per year, Reddit at an estimated $70 million per year, plus the Financial Times, Condé Nast, Hearst, Vox Media, The Atlantic, the Associated Press, Le Monde, and others. Google has its own Reddit deal estimated at $60 million per year granting real-time data API access. Anthropic has not publicly disclosed equivalent publisher licensing deals, and that undisclosed status is itself the practitioner-facing point. The corpora that fed these models, and that continue to refresh them, are not the same documents. Practitioners cannot know what any given provider has paid for and what it hasn’t.
The crawler infrastructure diverges next. OpenAI runs three separate bots: GPTBot for training, OAI-SearchBot for search indexing, and ChatGPT-User for user-initiated retrieval. Anthropic runs three of its own: ClaudeBot for training, Claude-SearchBot for search, and Claude-User for user-initiated retrieval. Perplexity runs PerplexityBot and Perplexity-User. Google introduced Google-Extended in September 2023 as the user-agent that controls whether Google can use a site’s content to train Gemini, separate entirely from the Googlebot that handles traditional search indexing. There is no single AI user-agent. Every provider requires a separate rule, and the rules don’t translate cleanly across providers because the bots don’t do equivalent jobs in equivalent ways.
The retrieval architectures diverge structurally. ChatGPT has historically used Bing’s index as its primary web search source, and that connection appears to still be primary, though OpenAI continues to build out additional infrastructure alongside it. Perplexity built its retrieval system on a Vespa-based pipeline that treats documents and sub-document chunks as first-class retrievable units. Google’s Gemini uses Google’s own index plus Knowledge Graph grounding. Claude uses Brave Search as a retrieval partner. Same query, four different retrieval systems, four different views of which sources exist and which sources are worth surfacing.
Then comes the alignment layer, which is where SEO had no equivalent at all. After a model is trained on its corpus, providers run post-training to shape how the model actually behaves: tone, refusal patterns, format, safety posture, what counts as a good answer. OpenAI’s primary approach has been RLHF, or Reinforcement Learning from Human Feedback, where human raters score model outputs and the model learns to produce highly rated responses. Anthropic developed Constitutional AI, which trains models to critique and revise their own outputs against a written set of principles. These methodologies produce demonstrably different behavior in the final products. The same retrieved content, fed into two models aligned by two methodologies, can yield two materially different responses about the same brand.
When one provider’s guidance demonstrably fails to port
The clearest single example of guidance that doesn’t port is llms.txt. Jeremy Howard of Answer.AI proposed the file in September 2024 as a markdown manifest, placed at a site’s root, that would guide LLMs to the most important content. The proposal got picked up across the SEO community. Yoast built a generator. Agencies added llms.txt creation to their service catalogs. Conference speakers declared it essential.
As of mid-2026, no major LLM provider has confirmed they consume the file. Not OpenAI. Not Anthropic. Not Google. Server-log analyses across hundreds of thousands of domains show major AI crawlers don’t routinely request /llms.txt at all. Google’s John Mueller publicly compared it to the deprecated meta keywords tag. Gary Illyes confirmed at Search Central Live in July 2025 that Google does not support llms.txt and is not planning to.
I’ve written about this elsewhere, so I won’t repeat the technicalities here. What matters for this argument is the structural lesson. Schema.org succeeded because three engines built it together and then enforced it together. llms.txt was proposed by one researcher, picked up by tooling vendors, and ignored by the platforms it was supposed to serve. The shared-standards model that gave SEO its portable guidance is not available to LLM practitioners at the same scale, because the platforms are not building the standards together. They are building their own pipelines.
The Gemini inversion
The cleanest illustration of how far guidance portability has degraded sits inside on…
When a brand stops appearing in ChatGPT, or when its share of voice in Perplexity drops by half over a quarter, the typical response from the marketing org is to write more content. Sometimes a lot more. The thinking goes that if AI systems aren’t surfacing the brand, the fix is to feed them more material to work with. That instinct is a misdiagnosis. It’s a retrieval-layer fix being applied to what is increasingly a different kind of problem entirely, and the cost shows up as wasted budget, missed quarters, and a creeping sense that the work isn’t connecting to the outcomes anymore.
The mistake is treating AI visibility as a single problem when it isn’t. There are three structurally different layers between your brand and the answer a user receives, each with its own failure modes, its own fixes, and increasingly its own organizational owner. Diagnose the wrong layer and the fix doesn’t land.
Where most of the conversation has been living
The first layer is retrieval. This is where the AI search optimization conversation has spent most of the last two years. The mechanics are familiar in shape if not in detail. When a model needs to answer a question grounded in real-world content, it pulls relevant material from external sources and uses that material to construct the response. The technical name is retrieval-augmented generation, or RAG, and the layer it operates on is the gateway between your content and the model’s output.
This is where crawlability, parseability, and chunk-friendliness do their work. If your content can’t be retrieved cleanly, nothing downstream matters. The visibility tracking platforms most marketing teams have evaluated this year measure outcomes that depend on this layer functioning, which is why they tend to reward the same disciplines that produced good results in classical search: structured content, schema markup, self-contained answers, clean technical implementation.
But retrieval has a structural limit, and Microsoft Research has been unusually direct about it. Plain RAG, in their words, struggles to connect the dots. It retrieves chunks of text that look relevant to the question, but it cannot reason about how those chunks relate to each other. When the answer requires synthesizing information across multiple sources, or when the question is broad enough that the right answer depends on understanding patterns across an entire dataset, retrieval alone breaks down. The model gets the chunks and has to guess at the relationships, and guessing is where hallucinations enter.
The discipline question this layer asks is straightforward. Can the model retrieve our content at all, and is it retrieving the right content for the right query? Most marketing teams have some version of this work in flight already, even if the specific tactics have shifted from classical SEO. But retrieval is only the gateway. Even when a model retrieves your content correctly, what it does with it depends on whether you exist as a recognized thing in the layer above.
Where entity recognition does the real work
The second layer is the relationship layer, and the dominant structure on it is the knowledge graph. The major search infrastructures all maintain one. Google’s Knowledge Graph, Microsoft’s Satori, and the open knowledge graph built on Wikidata and schema.org collectively define how your brand is represented as an entity, what category you sit in, and which other entities you’re connected to.
This is the layer that decides whether AI Overviews and large language model responses treat you as a recognized member of your category, or as one fuzzy candidate string among many. Brands that exist as clean, well-defined entities get cited consistently. Brands that exist as undifferentiated tokens scattered across the open web get pattern-matched against fifty other candidates and lose more often than they win.
Knowledge graphs have been around long enough that the discipline is reasonably mature. Schema markup on owned properties, consistent naming and identifiers across the open web, structured presence on the high-trust nodes like Wikidata entries and review platforms, and the slow accumulation of brand mentions in contexts that the graph treats as authoritative. This is where the unlinked brand mentions conversation lives, because consistent contextual mentions strengthen the entity even without a hyperlink attached. The fix at this layer is structural rather than volume-based. Writing more content does almost nothing if the entity definition underneath it is fuzzy.
The discipline question here is harder than the retrieval-layer question. Are we a clean, defensible entity in our category, or are we still being pattern-matched against fifty other candidate strings? A brand that can’t answer that question affirmatively is going to lose ground in AI search regardless of how much content it produces, because the second layer is where the model decides what your content is actually about.
The knowledge graph tells the model what your brand is. But increasingly, your brand has to function inside a third layer that most marketing teams haven’t met yet, where the model isn’t just understanding you, it’s being asked to reason about you on behalf of someone making a decision.
The layer enterprise companies are quietly building right now
The third layer is the context graph, and this one needs a careful introduction because most of the marketing conversation hasn’t reached it yet.
A context graph has the same structural shape as a knowledge graph, with entities, relationships, and typed connections, but it’s grounded differently. A knowledge graph models the world. It tells you what things are and how they relate in general. A context graph models a specific organization’s data, decisions, policies, and operational reality. The cleanest framing I’ve seen calls a knowledge graph the library and a context graph the operating manual written by the people who actually run the place. The library tells you what exists. The operating manual tells you what’s relevant, what’s authorized, and what to do about it right now. The library is read-only semantic infrastructure. The operating manual is a living operational layer that grows every time a business process executes.
What separates a context graph from anything that came before it is that governance lives inside the graph rather than alongside it. Policies, permissions, validity windows, and authorization rules are nodes the graph itself queries, not external documentation applied at the edges. When an agent retrieves something from a context graph, the result has already been filtered through what’s currently authorized, currently valid, and currently applicable. The graph is also continuously evolving, so what it knows about you this week is not necessarily what it knew last quarter. That’s where the word “governed” comes from when people in this space talk about governed retrieval. It isn’t a frame, but rather the architecture.
That architecture used to be invisible to anyone outside the organization that built it, which is why marketers haven’t had to think about it. That changed at Google Cloud Next ‘26, when Google introduced the Knowledge Catalog inside its new Agentic Data Cloud. Google’s own description of the product, written in their own first-party blog content, says the Knowledge Catalog constructs a unified, dynamic context graph of your entire business, enabling you to ground agents in all of your business data and semantics. That sentence is the moment the term left the data-engineering blogs and entered enterprise procurement vocabulary.
The reason this matters for marketing is that context graphs are what’s going to power the next generation of agents inside your enterprise customers. Gartner projects that 40 percent of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5 percent in 2025. Procurement agents, competitive intelligence agents, content st…
Search engines were designed to do several things at once: rank a field of options, route the user to one of them, and keep the human inside the decision so the engine never owned the choice. That last part was not an accident. It was the liability architecture. LLMs were built without any of it. They were built to answer the question directly, which is a different job entirely, and the design choices that follow from it change what visibility looks like, what risk looks like, and what the word ROI can honestly mean when the thing sending you traffic was never built to send traffic in the first place.
Two Systems, Two Jobs
A search engine’s job description is long. It crawls the web, indexes it, ranks a pool of candidate results against a query, presents them as a ranked list, and then waits for the human to make a click decision. The SERP itself has been drifting toward retention for years now, with galleries, rich snippets, answer boxes, local maps, video carousels, and AI Overviews all layering in features that keep the user on the page longer and route fewer of them to third-party sites. But the underlying contract was always the same. The engine offers options. The user selects one. The user owns the choice.
An LLM does not offer options. It produces an answer. The citation, when it appears, is not functioning as a routing instrument. It is closer to a grounding artifact produced by a retrieval pipeline, or in some framings a confidence hedge, or both at the same time. Whichever read you prefer, none of them describe a system designed to send traffic somewhere else. The system was designed to resolve the question in place.
That distinction sits beneath every metric conversation in this space. When practitioners ask what the LLM referral rate is, what the attributed traffic number looks like, what the click-through from an AI answer is, they are asking questions that assume a routing mechanism that is not actually part of the architecture. Whatever traffic does come through is a byproduct, not a design goal, and confusing the two is the first mistake in almost every conversation about AI visibility ROI.
The Liability Surface Moved
The human in the click decision was the SERP’s shield. If the link the user selected led somewhere harmful, misleading, or defamatory, the engine could point to the list of options and the user’s own agency in choosing one. The engine had not published the claim. It had surfaced ten candidate sources, the user had chosen one, and whatever happened next was not the engine’s editorial output. That is not a small feature. That is the reason Section 230 protections were structured the way they were, and why algorithmic ranking has traditionally been treated as something other than direct speech.
LLMs have no equivalent shield to stand behind. The system is producing the answer directly, in its own voice, without a field of options or a user-selected source. The liability surface that the SERP was designed to offload sits with the model producing the output, and the cases that have already moved through courts are starting to sketch the edges of that surface.
Walters v. OpenAI was dismissed on summary judgment in May 2025, and the decision leaned heavily on OpenAI’s disclaimers and a sophisticated reader who reasonably knew the chatbot could hallucinate. That reading protects general-purpose consumer chatbots in a very specific kind of case. It does not protect every product that uses a language model. In a separate matter, Air Canada was held liable for its customer service chatbot’s false statements about its own bereavement fare policy, because a customer could reasonably rely on an airline’s branded support agent for accurate information about that airline’s policies. Reasonable reliance is the key legal term, and the more specialized and authoritative the chatbot appears, the harder the disclaimer defense becomes to run.
The active litigation is still mapping the frontier. OpenAI is currently facing multiple lawsuits tied to allegations that ChatGPT drove users toward suicide or harmful delusions, several involving minors. The New York Times copyright case against OpenAI was allowed to proceed by a federal judge in March 2025, and Anthropic settled with book authors in August 2025 for a reported sum well into the billions. European GDPR complaints continue to move through Noyb. Battle v. Microsoft is still live. None of these outcomes are settled, and some will be dismissed on the same disclaimer grounds that resolved Walters. The point is not that LLM operators will lose every case. The point is that the liability surface now sits with the system producing the output, whether the individual plaintiff wins or loses, and every brand building against an LLM inherits some version of that surface when it uses the system’s output in its own customer-facing work.
The Denominator Problem
The most common argument against investing in AI visibility work sounds decisive until you look closely at what it is measuring. The argument runs roughly: ChatGPT and the others send a tiny sliver of referral traffic, somewhere in the low single digits of total inbound, so why reallocate budget toward a channel that barely moves the needle? Conductor’s research pegs the combined AI referral share at about 1% of publisher traffic. That number is real. At first read, it seems to close the ROI question cleanly.
It closes nothing. The problem is the denominator.
While the AI share of traffic holds roughly steady, the absolute volume of search-driven traffic has collapsed across most publisher categories. Similarweb data shows organic traffic to news publishers fell from about 2.3 billion monthly visits in mid-2024 to under 1.7 billion by May 2025, a loss of more than 600 million visits in under a year. Business Insider’s search traffic dropped 55% between April 2022 and April 2025, HuffPost lost roughly half of its search referrals, and The New York Times saw search’s share of its desktop and mobile traffic slide from 44% to 37%. Zero-click searches climbed from 56% to 69% between May 2024 and May 2025 as AI Overviews expanded across the SERP. A Reuters Institute survey of 280 media leaders in late 2025 found they expect another 43% decline on average over the next three years.
Read against that backdrop, a stable percentage share of a shrinking pie is not stable. It is a loss. The skeptics who point at the 1% number are measuring relative share of a traffic base that is contracting underneath them, and they are treating a falling absolute as if it were a steady state. The real question is not whether LLMs are sending meaningful traffic yet. The real question is whether the channel that used to send meaningful traffic is still doing what it used to do, and the answer is visibly no. The denominator is moving, and any ROI calculation anchored to the old denominator is a calculation of the previous environment, not the current one.
What The Billions Say
If the design-intent and liability and denominator arguments still leave room for doubt, the last place to look is revealed preference. What are the companies with the most complete internal data on user behavior actually doing with their capital?
The answer is unambiguous. The five largest US cloud and AI infrastructure providers have committed between 660 and 690 billion dollars in 2026 capital expenditure, nearly doubling 2025 levels. Alphabet alone is guiding to between 175 and 185 billion for 2026, more than doubling its 2025 spend of 91 billion. Microsoft, Amazon, Meta, and Oracle are all running similarly aggressive curves. The number that matters most, and that defuses the usual counter-argument, comes from Bank of America credit strategists who estimate AI capex will reach 94% of operating cash flows in 2025 and 2026, up from 76% in 2024.
That is not the shape of a defensive hedge. A hedge is a fraction of the cash flow, deployed to avoid being caught flat-footed if a competitor’s bet pays off. Companies d…
From the publisher's feed