How to get your software recommended by AI assistants

Which crawlers decide whether an assistant can see you

Separate the two kinds of crawler before you edit robots.txt. Blocking the training crawlers, GPTBot and ClaudeBot, has no documented effect on being cited. Blocking the search crawlers is a decision to leave those products.

One honest complication: blocking is leaky. A study of four million citations across 3,600 prompts found that sites blocking OAI-SearchBot were still cited 82.4 percent of the time, and sites blocking ChatGPT-User 70.6 percent. That fits the finding that most citations are served out of a search index rather than a live fetch. OpenAI's sentence is still true as written, and blocking remains an unreliable lever in both directions. Allow the search crawlers, then stop thinking about it.

Your pricing page, and the JavaScript split

"How much does X cost" is one of the most common buyer questions, and the answer is only as good as the page it came from. Two different audiences read that page. Indexing crawlers fetch your initial HTML and have no browser engine, so a price assembled client-side from an API call is a price they cannot see. User-triggered agent sessions do run a real browser, and their task success on web interactions roughly doubled over the last year, from about 30 percent to about 60 percent in repeated tests. A page can therefore be invisible to the index and workable for an agent at the same time, which is the worst of both worlds, because the index is what gets cited.

How much structure is worth: a controlled A/B of the same shop built human-first and then agent-ready, identical catalogue and prices, moved strict task success from 49.3 percent to 89.3 percent across 300 runs, with the largest gains in product detail extraction and comparison. Concretely: price in HTML text, no login, not excluded in robots.txt, server-rendered, with the currency, the unit and the billing period next to the number, and a date on the page. If you sell only by quote, publish a starting price and the drivers.

Answer-first question pages, and the limit of rewriting

Write one page per real buyer question, in the buyer's words, answer in the first two sentences, each answer self-contained because retrieval works on passages. Then stop optimising. The honest finding of the field is that rewriting for assistants is weak and sometimes harmful: a peer-reviewed benchmark concluded that "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking", while "traditional SEO strategies, those aiming to improve the ranking of the source in the LLM context, are significantly more effective". A 2026 critical survey of 45 GEO studies found that the widely quoted gains from the original 2023 GEO paper are "conditional on a source already being present in a fixed context" and "establish neither organic discoverability nor durable traffic effects".

Does structured data or llms.txt help?

Structured data earns its place for Google rich results and as a machine-readable copy of your facts: Organization, Product, Offer with a real price, FAQPage where the questions are visible. Google is explicit that it is not an AI requirement: "There's also no special schema.org structured data that you need to add." It is also policed. Marking up content readers cannot see, or emitting review markup for reviews you do not have, risks a structured data manual action and the loss of rich result eligibility. Markup mirrors the page, it does not make claims.

llms.txt is a proposal by Jeremy Howard from September 2024 for a markdown file telling agents how to use your site. It is a convention, not a standard: Chrome's Lighthouse audits for its presence and several documentation platforms generate it, but no assistant documents using it to decide who gets named. Thirty minutes of work, no expected traffic, and not something to buy as a service.

Why review gaming and invented statistics backfire

Buying reviews is a legal exposure. The US Federal Trade Commission adopted a rule in August 2024 banning fake and purchased reviews, undisclosed insider reviews, company-controlled "independent" review sites and the suppression of negative reviews. In the EU, rules applicable since 28 May 2022 require anyone displaying reviews to state whether and how they verify that reviewers are real customers, and submitting fake reviews is prohibited outright.

The subtler trap is invented numbers. "Add statistics" was the strongest single tactic in the original GEO paper, and when a 2026 study published GEO-rewritten pages to the live web, its faithfulness checker flagged all 80 of the Statistics Addition descriptions as numerically unfaithful. Adding true, sourced numbers is good practice. Adding plausible-sounding ones puts a claim in writing that contradicts every other source about you, and that mismatch is what gets a product dropped from an answer.

Evidence, contested, myth

Evidence: retrieval dominates content tweaking, 32 percent against 2.8 percent of explained variation; comparison listicles are the most cited format; most citations go to sites the subject brand does not own; AI citations come mostly from ordinary search indexes; brand size tiers visibility at roughly 73, 44 and 11 percent; agent-ready product and price markup more than doubles extraction success; crawler access is documented by every major vendor.

Contested: whether blocking a search crawler actually removes you, since the policy says yes and the measurement says 82.4 percent of blockers were still cited; whether question-and-answer structure helps, since Google says no special structure is needed while Bing recommends headings, tables and FAQ sections.

Myth: guaranteed positions in assistant answers; "GEO" rewrites as a reliable win, which the replication literature contradicts; keyword-stuffed llms.txt files; schema as a place to assert what the page does not say; buying reviews to impress a machine that reads several sources at once.

What we do, and what you can do this week

We publish comparisons that assistants retrieve, with dated prices and sources, and a free editorial listing is available whether you pay or not. A paid plan buys verified facts or a labelled sponsored placement, never a ranking position and never an editorial score, and every paid link carries rel="sponsored nofollow" as Google's spam policy requires. Thinking is published by TTA Technologies, which also builds LUXPOS, Salonify, BoxBee and DepotGest, so every page naming one carries a disclosure and the same published criteria.

Your week: allow the search crawlers, server-render your pricing page, make your name, address and phone identical everywhere, write three question pages, and work through the directories. Then read how to get your software listed in comparisons and send us three links at /en/get-listed.