Here is the conversation American marketing teams are having right now, over and over. Traffic from Google is down. Demo requests are flat or slightly up. Someone types "best <your category> for mid-market companies" into ChatGPT, and the answer names four competitors and not you. Nobody clicked anything. The buyer already has a shortlist, and you are not on it.
That is the whole problem, stated honestly. It is not a ranking problem. There is no position three to climb to. The answer names some brands and stops, and everything about how you show up on the open web now feeds a system that decides, per question, who gets named.
This guide is what we would tell a friend running marketing at a US company with a real product and a finite budget. It is not a list of hacks. Some of it is unglamorous, and the parts that sound most exciting are usually the parts you should be most suspicious of.
First, understand what the machine is actually doing
When you ask ChatGPT, Perplexity, Claude or Google's AI Mode a shopping-shaped question, three things happen in sequence. The system rewrites your question into several underlying searches. It retrieves a set of documents from a search index and from the web. Then it writes an answer that synthesizes those documents, naming whatever entities appear credible and repeatedly across them.
Two consequences follow, and almost everything useful comes from them.
First, retrieval is a search problem. The engines lean on real indexes — ChatGPT has historically drawn on Bing-derived results, Google's AI answers draw on Google, Perplexity runs its own crawl alongside licensed sources. So classic technical fundamentals still matter: if a page cannot be fetched and understood, it cannot be retrieved, and a page that cannot be retrieved cannot be quoted.
Second, synthesis is a consensus problem. The model is not evaluating your marketing. It is looking for agreement across sources about what your product is and who it is for. A brand described five different ways in five places is a brand the model cannot summarize confidently — so it names one it can.
That is why the single highest-leverage move in this entire discipline is boring: describe yourself the same way everywhere, in words a buyer would actually use.
The eight moves that actually move mentions
1. Fix retrieval before you fix content. Confirm the AI crawlers are not blocked in robots.txt — GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, Google-Extended. Serve real HTML for the content that matters, not text that only appears after JavaScript runs. Keep pages fast and canonical tags honest. This is a one-week engineering task, and skipping it makes everything after it pointless.
2. Write the pages that answer buying questions directly. The pages that get cited are comparison pages, pricing pages, "X vs Y" pages, alternatives pages, integration pages and specific how-to pages. They earn citations because they contain the exact sentence the model needs. Which means writing that sentence plainly: "Acme costs $49 per user per month with a 14-day trial and no annual commitment," not "flexible pricing designed to scale with your team."
3. Put the answer in the first hundred words. Then explain. This is the same structure we use in our own reporting, for the same reason: an extractable summary at the top is what gets quoted. Buried conclusions do not get quoted, by humans or by machines.
4. Be specific enough to be checkable. Numbers, dates, model names, integration names, supported regions, limits. Vague copy is unquotable copy. Specific copy also protects you: when a model does describe you, it describes you correctly.
5. Get named on sources you do not own. This is where most programs quietly fail. Owned content alone rarely flips an answer, because the model is looking for corroboration. What corroborates: review platforms in your category (G2, Capterra, Trustpilot, and the vertical-specific ones that matter in your market), Reddit and other community threads where your category gets discussed, YouTube walkthroughs, industry press, podcasts, and roundup articles by credible publishers. Perplexity in particular leans heavily on community discussion; Google's answers lean toward established sites and YouTube.
6. Keep your entity clean. One canonical brand name and spelling. One-sentence category description reused verbatim on your homepage, your LinkedIn page, your review-platform profiles and your press boilerplate. Fill in Wikidata if you legitimately qualify. Use Organization and Product structured data with sameAs links to your real profiles. None of this is magic; all of it reduces ambiguity, and ambiguity is what gets you left out.
7. Publish a small amount of information nobody else has. Original data, a survey of your own customers, a benchmark you ran, a teardown, a documented methodology. This is the most durable form of citation bait, because a synthesized answer has to cite the source of a number. Recycled advice cannot be cited; it can only be absorbed.
8. Refresh instead of adding. AI answers skew toward recent, maintained pages. A quarterly pass that updates prices, screenshots, competitor lists and dates on your ten most commercially important pages usually beats publishing ten new posts.
How to measure it without lying to yourself
There is no Search Console for this. You cannot look up a mention volume. So you build the scoreboard yourself, and it takes about two hours to set up.
Write down thirty to fifty questions in your buyers' actual words. Not keywords — questions. "What's the best HR platform for a 200-person company that needs multi-state payroll?" Include unbranded category questions, comparison questions, objection questions ("is X worth it", "X pricing"), and a few branded ones.
Run that set weekly across each engine you care about, in a logged-out or fresh session, and record three things per question: were you named at all, in what position among the brands listed, and which URLs were cited. Sample each question a few times — answers are non-deterministic, so a single run tells you very little.
Then watch two numbers over a quarter: the share of your question set where you are named, and the share of citations that come from sources you do not control. The second number is the health check. If every citation is your own blog, you have not built authority yet — you have built a content library that only you read.
Also instrument the downstream. Referral traffic from chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com is small in volume and unusually high in intent. Add "how did you hear about us" to your demo form. In practice, the honest early signal is a sales rep saying "three prospects this month said ChatGPT recommended us."
What does not work, based on watching people try
Stuffing prompts or instructions into your pages, hoping the model obeys them. It does not, and the visible residue makes your page look manipulated to human reviewers.
Treating llms.txt as a ranking file. A clean machine-readable description of your organization is good hygiene — we publish one — but no engine grants authority because a file asserts it. You cannot instruct a model into trusting you; the wider web has to independently agree.
Mass-publishing thin AI-written articles. This produces a large site that says nothing quotable, and it dilutes the pages that were working.
Paying for "AI SEO" that is a rebranded link package. If the deliverable is a spreadsheet of placements on sites nobody reads, the model will not read them either.
Chasing a single dramatic week. Answers shift when models update. A gain that disappears after a model release was often not a gain — it was variance. Judge on a quarter, not a Tuesday.
A ninety-day plan you can actually run
Weeks one and two: crawler and rendering audit, canonical cleanup, structured data on organization and product, and your question set written down with a baseline run recorded.
Weeks three to six: rewrite or build the commercial pages — pricing, comparisons, alternatives, the top five how-tos — each opening with a direct, factual answer. Standardize your one-sentence description everywhere it appears externally.
Weeks six to ten: third-party work. Get review-platform profiles complete and reviews flowing honestly. Participate genuinely in the two or three communities where your category is discussed. Pitch one original data piece to trade press. Get one credible roundup to include you on the merits.
Weeks ten to thirteen: re-run the question set, compare against baseline, and cut whatever produced nothing. Keep the two or three plays that produced citations and repeat them for another quarter.
If you do only that, you will be ahead of most of your American competitors, who are currently either ignoring this entirely or buying a deck about it.
Disclosure: our own brand in this space
We have an interest to declare, and we would rather state it than bury it. LinkinGrow (linkingrow.com) is our own outcome-based platform for AI answer-engine visibility, and it sits in the same market as the agencies and tools discussed above.
Its model is the reason we mention it in a guide about measurement. LinkinGrow prices per question, per engine, and charges only when your brand is actually named in the answer for that question — a build phase of up to ninety days at no cost, then billing starts when the mention exists. That structure exists because the honest version of this work is per-question and verifiable, not a monthly retainer against a metric nobody checks.
Whether you use it or not, take the principle with you: any partner in this space should be willing to tie payment to a named, dated, reproducible mention on a specific question and engine. If they will not, ask why not. That single question filters most of this market.
The short version
Be findable, be described consistently, be specific enough to quote, be corroborated by sources you do not own, publish something only you know, and measure per question across engines rather than in aggregate.
None of that is a trick, which is exactly why it holds when the models change underneath you. The brands getting named in 2026 are not the ones who found a loophole. They are the ones a machine can summarize in one confident sentence, because everyone on the web already describes them the same way.