01 The research
Language models named the leader and recommended a specialist
Being the best known business in your market is not what gets you recommended, and may be the reason you are not.
The brand a model recommended differed from the brand it named in 26 of the 42 categories tested.
A paper posted on 20 October tested 102 language models across 42 brand categories. In 26 of those 42 categories, the brand a model picked when asked for a recommendation was not the brand it named when asked who the leader was. The author’s own example is shoes: asked to name a running shoe brand the models say Nike, and asked to recommend one they say Brooks.
The pattern has a direction. The brands models named were market leaders, and the brands they recommended were, in the author’s words, mostly niche and specialist. The paper also reports that the models move together, settling on a shared recommendation rather than disagreeing.
The mechanism matters more than the shoes, because it is the part you can act on. Tracing three open training pipelines, the author finds the recommendation is attached during pretraining and then moved by the stage where a model is taught which answer a person would prefer, shifting the pick by 11 to 24 points. The data used in that stage contains almost no brands at all.
So the taste is not learned from facts about companies. The paper’s conclusion is that a recommendation depends less on what a model knows about a business than on whether that business fits a taste the model acquired elsewhere. The author is explicit about the limit, writing “We do not know exactly how these preferences arise,” and suspects the judge models used to rank answers during training.
For a surgical practice, that means the largest orthopedic group in the city is not automatically the one an assistant suggests, and a single-specialty practice is not automatically shut out. For a mergers and acquisitions adviser the same logic says a boutique can be picked over a national name, which is not how a ranked list of links has ever worked.
What to do about it
Describe what you specifically do, and who for, rather than asserting that you are the largest. On this evidence those are not the same signal, and only one of them is a claim a specialist can make honestly.
Source arXiv, “How Language Models Learn Taste: Brand Preference in LLMs,” 20 October 2026, 11 pages · arxiv.org/abs/2610.28272
02 The market data
An AI sales agent matched human staff on conversion
A machine answering your website chat converted as well as trained staff, at under two cents on the dollar per conversation.
The trial randomized 14.8 million visitors across 100 live websites in 14 industries over five months.
A preprint, which is a paper put out before peer review, was posted on 18 October. It reports a randomized controlled trial on 100 live websites across 14 industries, run from 1 May to 30 September 2026, in which 14.8 million visitors were randomly assigned to see no sales agent, a specialist one, or an agent built on a general-purpose frontier model. The specialist agent had been trained on 4.6 million consented, de-identified commercial conversations.
Against no agent at all, it changed revenue per visitor by 10.0 percent, and the paper puts the range at 5.1 to 14.9 percent. Against the general-purpose agent, among visitors who actually engaged, it changed conversion by 21.6 percent.
The number worth reading twice is the staffing one. On 20 of the sites that run staffed live chat, the agent’s engaged conversion was 11.26 percent against the humans’ 11.20 percent, which the paper is careful to call no worse rather than better, at 1.8 percent of the cost per conversation.
Now the part that decides how much of this to believe. This is a single-author preprint evaluating a specific named commercial product, which is a vendor study in a preprint’s clothes, and none of it has been peer reviewed. The trial design is genuinely strong and the safety limits were set in advance and reported, so the honest reading is that the method deserves attention and the result deserves replication by somebody with nothing to sell.
Forever Cited sells measurement in this field, so that last sentence applies to us as much as to anyone else quoting this paper, and we would rather say so than pass a 10 percent figure along as settled. For a charter operator or an information-security firm, where one inquiry is worth a great deal and the sales conversation is long, the practical question is not whether to install an agent but who answers for what it tells a customer about you.
What to do about it
If you are weighing up a chat agent, ask the vendor for the trial design rather than the headline number: how many sites, how visitors were assigned, and what the agent was measured against. This paper publishes all three, which is the bar, and most product pages do not clear it.
Source arXiv, “Selling in the Wild: Benchmark and Randomized Field Evidence for a Sales Agent Trained on Commercial Conversations,” 18 October 2026 · arxiv.org/abs/2610.25461
03 The risk
Researchers reached AI answers with one planted page
Where almost nothing has been written about a question your customers ask, one page by anyone at all can become the answer.
The planted material appeared in answers from ChatGPT, Gemini and Perplexity’s model 31 percent of the time.
Five researchers published a 23-page paper on 21 October, an extended version of work accepted at a security conference next year, describing what they call a data-void exploit. A data void is a question almost nobody has written about. The attack is to publish your own version of the answer into that gap and let an assistant’s web search find it.
They ran it against ChatGPT, Gemini and Sonar Pro, which is Perplexity’s model, varying the tone of the user’s question and three things the attacker controls. Across every condition tested, the planted content appeared in the assistant’s answer 31 percent of the time. The paper states that the attack can succeed with only a single attacker-controlled page.
A careful reader will ask whether a suspicious user is protected, and the answer is mostly no. Phrasing the question skeptically slightly reduced the success rate and slightly raised the chance the assistant attached a warning, and the authors report no evidence of a large effect from either the tone of the question or the choice of model. A ranked list of links lets you see where an answer came from and what the competing answers were, and the paper’s point is that a single synthesized answer takes both of those away.
One of the three attacker controls is uncomfortable reading for this industry. The researchers tested whether adding the content features thought to improve the chance an assistant cites a page made the attack work better, which means the techniques sold as visibility optimization were studied here as an attack tool. They are the same techniques either way, and what separates the two is whether what you publish about your business is true.
The defensive version of this is not a trick. If the honest answer to a question your buyers ask does not exist anywhere in writing, that gap is yours to fill before somebody else fills it. For a franchise advisory or an engineering practice, the questions most likely to be voids are the narrow technical ones only a handful of people in the country can answer.
What to do about it
List the five questions a serious buyer asks you on the phone, then search each one and see whether a credible written answer already exists. Where it does not, you are looking at the kind of gap this paper filled, and the page that fills it should be yours.
Source arXiv, “Data-Void Exploits in Web-Searching LLMs,” 21 October 2026, 23 pages, extended version of a paper to appear at the Network and Distributed System Security Symposium 2027 · arxiv.org/abs/2610.30365
04 The work
Researchers scored 18,796 work tasks for AI delegation
Only one task hour in forty can be handed to a machine completely, and those are the twelve-minute jobs.
Full delegation peaks at tasks that take a person about twelve minutes, and falls away for anything longer.
A 32-page paper posted on 22 October scored all 18,796 task statements in the United States government’s occupational database for whether an AI agent could take them over. The finding is in its first line: 35.2 percent of task time across world employment is at least partly delegable to AI agents in 2026, and only 2.5 percent can be handed over in full.
The second number is the one to hold onto. Partly delegable describes most of the working day; fully delegable describes one hour in forty. The gap between those two is the supervision somebody in your business still has to do.
The most useful detail in the paper is about size. Full delegation is rare and peaks for tasks that take a person about twelve minutes, then falls away for longer ones. So the work a machine can genuinely finish on its own is the short, bounded, repeated task, not the engagement.
This study is unusually well built for the genre. The scoring rubric was written and registered publicly before the scoring ran, the tasks were weighted by actual employment figures across 146 countries, and a person who graded 300 of the same tasks without seeing the models’ answers landed on the same rating the overwhelming majority of the time. The index, the prompts and the code were all released.
For an engineering practice, the twelve-minute finding points at the drawing check, the submittal log and the standard calculation, not the design judgment. For a charter operator it is the quote turnaround and the document chase rather than the decision to fly, and working out which of your tasks are twelve minutes long is a more useful audit than asking whether AI will replace your profession.
What to do about it
Make a list of the tasks in your week that take about twelve minutes and keep recurring. That is the list this paper says a machine can actually finish, and it is a far better place to start than a general question about whether your profession is exposed.
Source arXiv, “The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries,” 22 October 2026, 32 pages, preregistered · arxiv.org/abs/2610.30752
05 The money
Nearly a third of buyers skipped a software purchase
If you sell software, some of your renewals are now quietly being rebuilt in-house by the customer.
The figure is self-reported, and the authors say plainly that it does not show how much spending actually moved.
A 37-page paper posted on 21 October asks whether AI coding agents change the decision to buy software or build it. In an organizational survey, it reports that 32 percent of respondents had gone without at least one software purchase because their own organization could build the function with coding agents.
The capability trend behind that is measurable. Since 2023, the length of task that leading coding agents can complete with 50 percent success, measured in human working time, has doubled roughly every 129 days. A search by the authors turned up 575 possible open-source alternatives to a selected set of commercial products.
The authors are notably careful, and that is the reason to trust the rest. They write that the self-reports do not establish the spending displaced or the reliability of the resulting tools, that some of those 575 alternatives still need independent checking, and that the 12 applications they built themselves remain early versions with unverified completeness and reliability.
Their own summary of the projection is that it is “an extrapolation, not a forecast.” The market evidence they put beside it cuts both ways: purchased software gained share of United States investment and tracked commercial installations grew, at the same time as a third of surveyed buyers said they skipped a purchase.
For a software company selling to professional buyers, the warning here is not collapsing demand but a quieter change in who is at the table: the customer who knows the work well enough to specify the tool is now sometimes able to build it. For an information-security firm, that same capability is the reason an unreviewed internal tool turns up on your network with nobody’s name against it.
What to do about it
If you sell software, ask your next renewal conversation what the customer has already built for themselves this year. If you buy it, write down who maintains the thing you build instead, because the paper’s own caveat is that reliability is the part nobody has measured yet.
Source arXiv, “The Customer Is the Builder: How AI Coding Agents Are Rewriting the Economics of Software,” 21 October 2026, 37 pages · arxiv.org/abs/2610.30307
