B2LLM
B2LLM is business to LLM: the channel between an organisation and the large language models that now sit between it and its audience. We have had B2B and B2C for as long as anyone now working can remember. B2LLM names the third channel, the one that opened when people stopped clicking through to websites and started asking a model instead.
Where it came from
I proposed the term on stage at Reuters Pharma in April 2026, on a panel asking what declining web traffic means for engaging healthcare professionals. The panel had just heard the numbers: the moderator’s firm, Syneos Health, had measured clinicians’ regular use of language models jumping from roughly 70 to 96 per cent in a year, while flagship medical websites lost around 30 per cent of their traffic. Here is the moment, from the recording:
We’ve always got B2B and B2C. And I haven’t thought of this yet, but it’s maybe B2LLM: in the sense that we are putting stuff into the world for a whole host of companies and proxies to be the messenger of our content and our messages. And how we do that, we need to learn, we need to share, we need to test, we need to experiment.
Reuters Events later wrote up the panel, naming B2LLM and attributing it to me; their summary is itself AI-generated, which rather proves the point.
The argument
We should be optimising for a layer between us and our audience: LLMs. For twenty years the discipline was being found, first by search engines, then by feeds. My doctorate put a name to what finished that era: the shift from a deferrer society to a referrer society, in which people no longer simply accept what authority tells them; they check, compare, and increasingly ask an AI. When the answer comes from a model, the model is the reader you have to reach first. Not instead of the human; ahead of the human. The model is the messenger of your message, and it will carry someone’s version of the facts whether you show up or not.
Done honestly, optimising for that layer is not a dark art. It is the same instinct that has driven all my work in health communication: go where people actually are, and make credible information legible there. Over the last century this pattern has persisted. The platform used to be printed media, then radio, then television, then social media. Now it is the model itself. The speed and scale of this transformation in how information is disseminated is unprecedented. If the careful people opt out of this channel, the models get furnished by whoever optimises hardest, not by whoever is most correct.
Who it is for
The panel was a pharma room, but the concept applies far beyond the life sciences. B2LLM is for anyone whose audience now asks an AI before it asks them: publishers, educators, clinicians, brands, institutions. If you are building for the AI-search era, you are already in the B2LLM business, whether you have named the channel or not.
What B2LLM does not claim
Naming a channel is not a licence to abuse it, so let me be precise about the edges.
- It does not mean writing solely for machines. The model is the messenger; the eventual reader is still a person. But attrition from the answer back to the source is high. With many LLMs writing to please, critical appraisal skills are ever more critical.
- It does not replace communicating with humans. It sits ahead of them, not instead of them.
- It does not guarantee that a model will cite you or repeat you accurately. Field Test 001, this site’s before-measurement, shows the opposite is common: the models missed most of my record, cited pages I do not own, and in one case credited me with a colleague’s work.
- It does not justify flooding or manipulating answer engines. Done honestly or not at all.
- It is a proposed communication model, my coinage, not an externally validated discipline.
The claims, classed
I sort what I publish by how strong the claim is, so here is this page graded by my own rubric. The classes are Observation, Hypothesis, Coined term, Internal result and Externally validated finding; each claim names its evidence and where that evidence comes from.
The ledger
B2LLM itself. Coined term. Evidence: the timestamped recording of the Reuters Pharma panel, April 2026, where the words were said on stage (29 minutes 51 seconds in). Source: my own archive recording of the event, owner-supplied, primary. Also published: Reuters Events wrote up the panel in May 2026, naming the idea and attributing it to me. That is third-party corroboration that the term was coined on that stage; it is not independent validation of the concept, which remains my proposal.
Audiences increasingly consult language models instead of clicking results. Observation. Evidence: the panel discussion itself, and the usage my co-panellists described from clinical practice. Source: owner-supplied recording of the event; consistent with the moderator’s figures below, which I have not independently verified.
Clinicians’ regular LLM use jumped from roughly 70 to 96 per cent in a year, while flagship medical websites lost around 30 per cent of their traffic. The moderator’s on-stage account of research by his firm, Syneos Health. Evidence: quoted from the panel recording. Source: a third party’s claim inside an owner-supplied recording; not independently verified, which is why this page attributes it rather than asserts it.
The shift from a deferrer society to a referrer society. Internal result, institutionally examined. Evidence: my MD thesis (Hull York Medical School, 2025), an IPA study of 40 participants, self-hosted in full with its abstract at the research page. Source: owner-supplied document; the degree award is the institutional check, not external replication.
Asked cold about a specific person, today’s models answer from borrowed ground, miss most of the record, and sometimes misattribute. Observation. Evidence: Field Test 001, in which the same question about me was put to 12 AI sessions across 4 vendors before launch; the median cold session named 3.5 of 8 core facts, the 87 citations resolved to 27 distinct sites with almost all the weight on pages I do not own and my own domain appearing once, as a holding page, and one cold session credited me with a colleague’s work. Source: owner-conducted dataset, transcripts preserved verbatim as primary evidence.
Organisations should treat the model layer as a distinct channel and optimise for it. Hypothesis. Evidence: the argument on this page and the live experiment this site runs on itself, whose before-measurement is now published as Field Test 001 (12 AI sessions across 4 vendors, scored against this site’s own answer key). Source: my own proposed model and an owner-conducted dataset, not yet validated; the after-measurement, run once the site is live, is what would begin to validate it, and this page will say so until it does.
Watch and read
- The panel this term was coined on: What Does Declining Web Traffic Mean for HCP Engagement?, Reuters Pharma, April 2026
- The panel written up: Navigate AI-driven engagement: transform declining web traffic into direct HCP connections, Reuters Events, May 2026. Their summary is itself AI-generated, which is the argument of this page in miniature.
- The reflection I wrote when the idea earned its own page: B2LLM: when the reader is a language model
- The research underneath it: the deferrer to referrer shift, my MD at Hull York Medical School
- The working experiment: this site is built to be read by machines too
- B2LLM, measured: Field Test 001, the pre-launch baseline, the same question about me put to 12 AI sessions across 4 vendors before this site went live