Written for the machines

Written for the machines. Generative engine optimisation is a legal, well-capitalised industry that sells influence over what AI systems tell you. The academic evidence says the techniques work. Fraudsters and blue-chip PR firms worked that out before most marketers did. Ask Perplexity for the official Emirates reservations number and, for a period late last year, you could have been given a scam call centre. The system described it as a booking and upgrades hotline. Google's AI Overviews fell for the same campaign, returning fraudulent Emirates numbers alongside step-by-step booking instructions. The security firm Aurascape, which documented the operation in December, believes it to be the first real-world campaign built specifically to make AI assistants recommend fraudulent contact details. The operators did not hack anything. They seeded structured question-and-answer text across compromised government and university sites, WordPress blogs, YouTube descriptions and Yelp reviews, formatted for easy machine extraction, and waited for the models to come and collect it. The technique they used has a name, an academic literature, a trade press and, since February, a company built around it valued at a billion dollars. What generative engine optimisation is The term arrived in a paper. In November 2023, six researchers led by Pranjal Aggarwal published GEO: Generative Engine Optimization, later accepted to the ACM's KDD conference, asking what becomes of content creators when people stop reading ranked lists of links and start reading generated answers. The authors built a benchmark of diverse queries, tested a range of editorial interventions, and reported that their methods could boost a source's visibility inside a generated response by up to 40 per cent. The tactics that worked best were unglamorous: adding statistics, adding quotations from named people, adding citations to outside authorities. The mechanism is retrieval-augmented generation. Training data goes stale, so modern systems search the live web when you ask a question and compose an answer from what comes back. The open internet is therefore an input to the answer, and not merely a historical influence on the model's weights. That distinction separates two very different problems. Poisoning training data requires access to the training pipeline. Influencing retrieval requires a website and an understanding of what the retrieval system finds convincing. The first is hard. The second is a marketing service with a price list. Search engine optimisation applied the same logic to a different ranking function and grew into an industry worth more than $80bn a year. GEO inherits its practitioners and most of its tradecraft. What changes is the prize. SEO competed for a slot in a list the reader could inspect. GEO competes to become a sentence in an answer, with provenance reduced to a footnote or dropped entirely. Does any of it work It works, and less precisely than the people selling it imply. The strongest demonstration of deliberate manipulation comes from Harvard. Aounon Kumar and Himabindu Lakkaraju showed that inserting an optimised strategic text sequence into a product page can lift that product to the top of an AI shopping recommendation. Working with a catalogue of fictitious coffee machines, they took products that rarely appeared at all and made them the first suggestion. The sequences were generated adversarially, using the same gradient methods that produce jailbreaks, so this sits closer to an attack than to marketing. It establishes the principle that the ordering inside a generated answer is manipulable by whoever controls the retrieved text. The most useful scepticism comes from inside the marketing industry. Rand Fishkin of SparkToro, working with Patrick O'Donnell of Gumshoe, ran twelve prompts through ChatGPT, Claude and Google's AI systems 2,961 times using 600 volunteers. Two responses to the same prompt returned an identical list of brands less than one time in a hundred. Fishkin concluded that tracking a brand's rank in AI answers is meaningless and that a good deal of what the industry sells amounts to measuring randomness. The finding that survived was more interesting. The pool of brands a model draws from proved reasonably stable, with the leaders in each category appearing in 55 to 77 per cent of responses regardless of phrasing. That is the shape of the thing. Position cannot be bought. The odds of being in the pool at all can be shifted, which for anyone trying to place an idea rather than a product is the entire objective. The marketplace The commercial figures are soft and moving quickly. One market research firm put the global GEO services market at roughly $1bn in 2025, a number best treated as an early guess at a category nobody has agreed how to define. The capital is firmer. In February the New York platform Profound raised $96m at a $1bn valuation in a round led by Lightspeed, reaching about $155m in eighteen months. Three months earlier Adobe agreed to buy Semrush for $1.9bn in cash, framing the deal explicitly as a way to sell GEO alongside SEO. Demos surveyed fifty firms openly advertising these services, fifteen of them British, and found three broad offerings. Most optimise a client's own site for AI crawlers. Around a third create fresh content and place it on third-party industry and review sites, on the reasoning that models weight signals of independence. A smaller number work directly on the platforms that models cite most heavily. That last group matters out of proportion to its size, because Reddit and Wikipedia sit near the top of the citation tables for the major systems, and both are open to anyone with patience and an account. The reputable end of the trade Quietly editing Wikipedia for paying clients is an old business with a new payoff. In January the Bureau of Investigative Journalism reported that the London firm Portland Communications had commissioned favourable edits on behalf of governments and billionaires, using subcontractors to avoid detection. Six former employees described Wikipedia edits as a routine request from Qatar between 2013 and 2024, with pages covering the country's human rights record around stadium construction among the targets. Other reported edits touched a Gates Foundation agricultural programme and a dispute involving Libya's sovereign wealth fund. Portland, founded by Tim Allan, now Sir Keir Starmer's director of communications, says editing Wikipedia was never endorsed by the firm, was never widespread among employees and is not done by current staff. Former Portland employees are the closest thing this field has to whistleblowers, and one told the Bureau that the internal question was never whether to stop, only how to carry on without being caught. The Chartered Institute of Public Relations, whose code treats the practice as unethical, raised concerns about the firm afterwards. None of that had anything to do with AI when it started. It does now, because a Wikipedia article is no longer read only by people who visit Wikipedia. There is a reasonable defence of all this. Businesses have always tried to be findable, models genuinely do reward clarity and sourcing, and much of what GEO firms sell is indistinguishable from writing well. Demos, whose survey this draws on, is careful to say it accuses none of the companies in its sample of anything. The difficulty is structural rather than individual. Of the fifty firms surveyed, exactly one had a public policy addressing misuse of its services, and of the five major model providers reviewed, only Google has published anything explicitly addressing the practice. A capability is being built quickly, under strong commercial incentives, with almost no thought given to who else might want it. Fraudsters have already answered that question. Part II turns to a state that answered it too. --- Disclosure: Anthropic's Claude was among the models tested in the SparkToro research described above, and Claude was used in the research and drafting of this piece. Sources: Aurascape; arXiv and ACM SIGKDD; SparkToro; the Bureau of Investigative Journalism; Demos; Fortune; Adobe; the Chartered Institute of Public Relations.