How to make your science visible to AI search engines

Your buyers now ask AI engines before they ask Google. This is the playbook we run for life science clients: how engines choose what to cite, a structured FAQ strategy, the off-site placements that earn citations, and the KPIs that prove it works.

________________________________________________________________________________________________________________

Your buyers now ask ChatGPT before they ask Google

In March, a European biologics CDMO asked us a simple question: when a biotech asks ChatGPT for a shortlist of CDMOs, do we show up? We ran 18 real buyer prompts across four AI engines. The company appeared in 21% of answers. ChatGPT named it on 2 of 18 questions. Six months later, the same 18 prompts return it in 38% of answers, and on ChatGPT specifically it now edges past its closest peer.

That is the shift every life science brand is living through. A process development director evaluating partners no longer starts with ten blue links. She asks an AI engine for a shortlist, reads one synthesized answer with a handful of citations, and clicks on one or two of them, if at all. In that answer, your company is either named or it does not exist.

The traffic numbers still look small. AI engines sent 0.32% of all website visits in 2026 against 42.75% for organic search, according to SE Ranking's study of 101,574 websites. The same study shows that share has grown 16x in under two years, that ChatGPT alone accounts for three quarters of it, and that AI-referred visitors stay on site 68% longer than organic visitors. Small volume, high intent, and compounding fast. Waiting for the curve to flatten is not a strategy.

This article is the playbook we now run for life science clients: how AI engines actually choose what to cite, a structured FAQ strategy that gets your science quoted, the off-site placements engines favour, the technical hygiene that makes your pages machine-readable, and the KPIs that prove it is working. It ends with a prioritized action list you can start on this week.

How AI engines decide what to cite

Answer engine optimization (AEO) is the work of getting your pages, and your brand name, into the answers that ChatGPT, Perplexity, Gemini, Claude and Google's AI Overviews generate. It builds on SEO. It does not replace it. But the mechanics differ from classic ranking in three ways that change what you should do.

First, engines fan the question out. Google describes AI Mode as "breaking down your question into subtopics and issuing a multitude of queries simultaneously". A prompt like "which European CDMOs have FDA and EMA inspection experience?" becomes six or eight sub-queries: inspection history, 483 records, EU GMP certificates, capacity, cost. Each sub-query retrieves passages, not pages. Your content wins when one passage answers one sub-question cleanly.

Second, retrieval runs through search indexes. Ahrefs analysed 1.4 million ChatGPT prompts and found that URLs pulled from web search were cited far more frequently than anything surfaced from news, Reddit or YouTube (an 88% citation rate versus 12% or less). The strongest on-page predictor was how closely the page title matched the fanned-out query. Backlinks, word count and heading structure showed no meaningful correlation once the page was retrievable. If you are not indexed, and your titles do not say what the page answers, nothing else matters.

Third, every engine has its own habits. We saw them clearly across 18 CDMO buyer prompts over six months:

Semrush's 13-week study of 230,000 prompts confirms the pattern at scale: Perplexity and AI Mode lean on Reddit, LinkedIn and YouTube, ChatGPT's citation mix swings month to month, and LinkedIn is the one source rising on every platform.

The practical conclusion: on-site structure gets you retrieved, off-site presence gets you trusted. You need both, and you need them in that order.

Step 1: Build a structured FAQ strategy around real buyer questions

Most life science websites are written as brochures: capabilities, facilities, a leadership page. AI engines do not retrieve brochures. They retrieve answers to questions. A structured FAQ strategy turns what your scientists already know into passages an engine can lift verbatim.

Mine the questions before you write a word. Start with the questions your buyers ask a human: SME interviews, sales call transcripts, RFP checklists, conference booth conversations. Add the queries Google Search Console already shows you, the People Also Ask boxes on your category terms, and the follow-up questions Perplexity suggests under its answers. Then ask the engines directly: "What would a biologics development director ask before shortlisting a CDMO?" For our CDMO client, this produced a fixed set of 18 prompts, six each at awareness, consideration and decision stage. That set became both the content brief and the measurement baseline.

One question, one answer, one place. Write the question as the heading, in the buyer's words, and answer it completely in the first 40 to 60 words. Then add the evidence: the number, the named expert, the regulatory reference. Every paragraph should survive on its own, because that is how it will be quoted. The Princeton GEO study found that adding statistics, quotations and cited sources to a page lifted its visibility in generative answers by up to 40%. Vague copy gets paraphrased. Specific copy gets cited.

Match the title to the question. Ahrefs found the tightest predictor of a ChatGPT citation was semantic similarity between the page title and the fanned-out query, and that natural-language URLs were cited 89.8% of the time versus 81.1% for opaque slugs. So /faq/can-one-cdmo-cover-clinical-and-commercial-supply beats /resources/faq-2.

Put the FAQ where the question lives. Decision-stage questions ("What is your batch success rate?") belong on the service page they concern, with a short answer visible on the page. Category questions ("How do you compare tech transfer capability across CDMOs?") earn a standalone page each. For our client, standalone question pages reached Google positions 3 to 7 within a quarter of publication, and the first SME interview article landed at position 2.8 with an AI Overview inside five days. One consideration-stage prompt about end-to-end programme support went from 25% to 42% AI visibility the month the matching case-study content went live.

Mark it up, but do not rely on the markup. Google retired FAQ rich results in May 2026; FAQPage remains a valid schema type and costs nothing to keep accurate. Ship it alongside Organization schema and Article schema with author, datePublished and dateModified. The visible question-and-answer formatting on the page does the measurable work. The JSON-LD confirms it.

Keep it fresh. Across 17 million citations, Ahrefs found AI assistants cite content 26% fresher than Google's organic results, and ChatGPT and Perplexity order citations newest first. Review each FAQ quarterly, update the numbers, and update the dateModified stamp when you do.

Step 2: Earn the off-site placements engines already trust

On-site work gets you retrieved. It does not get you recommended. When we asked the engines which European CDMOs to shortlist, they did not build the list from company websites. They built it from third-party lists, directories, LinkedIn articles and trade press, then checked the company sites for detail. If you are absent from the sources an engine leans on, you are absent from the shortlist.

The data backs this at scale. Ahrefs studied 75,000 brands and found the strongest correlate of AI visibility was YouTube mentions (0.74), followed by branded web mentions (0.66 to 0.71). Backlinks showed almost no relationship. Mentions, not links, are the currency.

For our client, three weeks of citation tracking made the priority list obvious:

One warning from the same data. The single Reddit citation for the client concerned an unrelated account with a similar name. Engines confuse entities with similar names, and Claude in particular answers from what it has memorized. Use one exact brand descriptor everywhere, "Company, European biologics CDMO", in press releases, directory profiles, LinkedIn bios and schema. Consistency is how a model learns who you are.

Step 3: Make the science machine-readable

The biggest single jump we measured came from the least glamorous work. In March, the client's site had 21 pages without a title tag, including the homepage and all five service pages, 34 pages without a meta description, 198 images without alt text, and no schema of any kind. Seven weeks after we rewrote the title tags to match the buyer questions, shipped entity schema and rewrote the copy, four-engine AI visibility went from 21% to 44%. We did not isolate the alt text or the meta descriptions, and we do not credit them. The titles, the schema and the copy did the work. None of it was novel. All of it was missing.

Let the answering bots in. AI companies run two kinds of crawlers: training bots (GPTBot, ClaudeBot, Google-Extended) and answering bots that fetch pages when a user asks a question (OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot). Cloudflare's August 2026 robots.txt snapshot shows publishers now block training bots roughly 2:1 and allow answering bots. For a B2B brand that wants to be recommended, block neither. ChatGPT answered 14 of our 18 prompts without searching at all. What a model memorized about you in training is what it says when it does not look you up.

Serve HTML, not JavaScript. Most AI crawlers still do not execute JavaScript. If your capability tables, FAQ accordions or publication lists render client-side, the engines see an empty div. Server-render them, or make sure the text is in the initial HTML.

Get indexed everywhere, not just Google. Retrieval runs through search indexes, and ChatGPT's runs largely through Bing. Verify the site in Bing Webmaster Tools, submit the sitemap, and switch on IndexNow so new FAQ pages reach the index the day they publish. Check that regulatory PDFs, posters and white papers you want cited are indexable HTML, or at least have an HTML summary page.

Ship the schema that describes an entity. Organization schema with a one-sentence description, founding date, locations and sameAs links to LinkedIn, Crunchbase and Wikidata. Person schema for every SME author with credentials and affiliation. Article schema with dates. FAQPage on the question pages. For medtech and diagnostics, the MedicalDevice and MedicalTest types exist; use them with the same care you apply to claims in labeling. The client's homepage became the most-cited page across all AI answers (151 citations in three weeks) because it now states in its first line exactly what the company is, for whom, and where.

llms.txt: ship it in an hour, then forget it. We published one for the client and visibility spiked the same week. We also shipped schema, titles and rewritten copy in the same window, so we cannot credit the file. Independent log data says we should not: AI crawlers requested llms.txt in 0.1% of visits, Google states it does not use it, and SE Ranking found no correlation with citations across 300,000 domains. It costs nothing and hurts nothing. Do not build a roadmap around it.

Show who is talking. Named authors with PhD or MD credentials, an author page per SME, visible publication and update dates, and links out to primary sources (FDA guidance, EMA guidelines, peer-reviewed papers). In this category, Gemini and AI Overviews cite regulators and credentialed sources first, and the humans who read the citation weigh credentials too.

Measure it like a scientist: the KPIs that prove AEO works

AEO has a measurement problem. Referral traffic is tiny, answers change between runs, and no tool sees inside a model. The fix is the same one you would apply to any noisy assay: a fixed protocol, repeated runs, a baseline, and controls.

Our protocol for the client was deliberately simple. 18 prompts, word for word identical every month. Four engines. Each prompt asked three times per engine, so 216 data points per batch. Score: mentions divided by total runs (18 x 4 x 3 = 216), so a brand named on every run of every engine scores 100%. That produced a clean baseline (21.3% in March) and a defensible trend (38.0% in September) that a board can read. When we moved to an automated daily tracker in September, we kept the 18 prompts identical and declared a new baseline, because a weighted score is not comparable to a yes/no count. The 38% is the last manual batch; the automated tool's first three weeks put the client at 17% on its own scale. State your series changes. Boards remember the number you showed them last quarter.

Track these, in this order of importance:

Three lessons on reading the numbers. First, per-engine scores swing hard: ChatGPT swung from 56% in May to 13% in June and back to 48% in September, while the four-engine average stayed between 29% and 44%. Report the average and the trend, and treat any single-engine reading as weather. Second, set up the GA4 channel before you start. Define a custom channel group matching chatgpt.com, perplexity.ai, gemini.google.com, claude.ai and copilot.microsoft.com as source, and expect undercounting, because a share of AI-referred visits arrives with no referrer and lands in Direct. Third, tag lead source in the CRM on day one. Lead-source tagging was not in place for the client until month five, and no MQL was attributable to organic or AI for the first half year. The visibility story was strong. The pipeline story was blank. Do not repeat that.

On expectations: the technical layer moves the numbers within two months. Content and off-site work take three to six months to register. Judge the program at twelve.

What six months with one CDMO taught us

The honest trajectory on the same 18 prompts: 21% in March; 44% in May, seven weeks after the on-site work; 29% in June when ChatGPT and Gemini fell back; 38% in September as the content landed. ChatGPT alone went 11% to 56% to 13% to 48%. Non-brand Google keywords went from 255 to 439. Zero of it came from a trick. Here is what moved the needle, and what did not.

What worked. Title tags, entity schema and a copy rewrite around the buyer questions produced the first jump, and ChatGPT and Gemini reacted hardest. SME interview content produced the second wave: one consideration-stage prompt moved from 25% to 42% the month its matching article went live, and the September batch recovered to 38% as content and structured data landed together. Splitting one long article into two focused pieces published faster and ranked faster. We interviewed one expert a month and turned each interview into two or three single-question articles. That cadence, not any single piece, is the asset.

What did not work. Technical work alone did nothing for Perplexity, which wanted off-site mentions we had not yet earned. Targeting zero-volume "timeline" keywords produced nothing measurable. A marketing platform's built-in AI tracker covered three engines with placeholder prompts, so we stayed with our own protocol until a tool ran our exact prompts across all five engines. And the June dip taught us never to report a single batch as a trend.

Being retrieved is not being clicked. Seventeen commercial queries earned 1,513 Google impressions and four clicks in six months. Two of them sat at position 2 to 3 with zero clicks. The pages answered the question in the snippet and gave the reader no reason to visit. Every answer-first page needs a next step: a data sheet, a facility tour, a comparison, a person to talk to.

Your own testing pollutes your data. Several of the top non-brand queries in Search Console turned out to be our numbered test prompts, at 68 to 90 impressions each. Run prompt tests from a clean environment and filter them out of every report. We now do.

The bottleneck was never the writing. Nine articles were drafted in twelve weeks. Three went live. SME sign-off took 35 to 51 days per article against the three days planned. AEO compounds with publishing cadence, so a review SLA belongs in the plan on day one, with a named reviewer and a calendar slot, not a hope.

Off-site is where the next gains are. The topics engines score the client lowest on (regulatory inspection credentials, cost comparison, development timelines, a buyer's guide) and the eight directories that cite competitors but not the client are the Q4 plan. That is the pattern to expect: months one to three are on-site, months four to twelve are content and off-site authority.

The prioritized action list

Ordered by impact per hour of effort. Each block builds on the one before it.

Frequently asked questions

What is answer engine optimization (AEO)?

AEO is the work of getting your pages and your brand name into the answers that ChatGPT, Perplexity, Gemini, Claude and Google's AI Overviews generate. It builds on SEO. Engines retrieve passages, weigh third-party sources and answer from what they memorized.

How long does AEO take to show results?

In our client work, the technical layer moved AI visibility within two months. Content and off-site work took three to six months to register. Judge the programme at twelve months.

How do you measure AI search visibility?

Run a fixed set of buyer prompts every month, word for word identical, three times on each engine. Score the share of runs that name your brand. Our CDMO client went from 21% to 38% on 18 prompts in six months.

Does llms.txt improve AI visibility?

The evidence says no. AI crawlers requested llms.txt in 0.1% of visits, Google states it does not use it, and SE Ranking found no correlation with citations across 300,000 domains. It takes an hour to ship and does no harm.

Do backlinks matter for AI search?

Much less than mentions. Across 75,000 brands, Ahrefs found YouTube mentions and branded web mentions correlated most with AI visibility. Backlinks showed almost no relationship.

Can you buy your way into ChatGPT answers?

No. ChatGPT Ads run as a separate system. A sponsored card sits below the answer and does not change how often ChatGPT cites your pages.

Start with the question, not the tool

New AEO tools launch every week, and every one claims a secret. None of them change the mechanics: engines fan the question out, retrieve passages from search indexes, weigh what trusted third parties say about you, and remember what they have read. A life science brand wins by answering the buyer's questions in plain, specific, attributed passages, by being present on the lists and platforms the engines already trust, and by measuring the same 18 questions every month until the trend is undeniable.

Most of your competitors have not started. In our client's category, only two of seven peers had any FAQ content in March, and none answered the buyer questions in a form an engine can quote. The gap is open. It will not stay open.

Find out what the engines say about your science today. Run the 18-prompt baseline yourself with the protocol above, or send us your top buyer questions and we will run it, benchmark you against five peers, and hand you the list of pages and placements that will move the number first.

Sources

• SE Ranking, AI traffic research study: 101,574 websites, January 2025 to April 2026
• Semrush, The most-cited domains in AI: a 3-month study of 230,000 prompts
• Ahrefs, Why ChatGPT cites one page over another: 1.4 million prompts
• Ahrefs, Do AI assistants prefer to cite fresh content? 17 million citations
• Ahrefs, AI brand visibility correlations across 75,000 brands
• Aleyda Solis, Google AI Mode's query fan-out technique
• Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024
• OtterlyAI, LinkedIn AI search citations study: 1.3 million citations, January to June 2026
• Search Engine Land, AI search engines cite Reddit, YouTube and LinkedIn most (Peec AI, 30 million sources)
• Passionfruit, What changed when Google dropped FAQ rich results, May 2026
• HybridRanking, llms.txt one year later: who is actually reading it
• HybridRanking, Most AI crawlers still do not render JavaScript in 2026
• TechnologyChecker, robots.txt across Cloudflare's network: publishers block training bots and allow answering bots, September 2026
• Analytics Mania, How to track and report AI traffic in Google Analytics 4

Client figures come from Ooho's monthly AI visibility tracking, Google Search Console and GA4 for a European biologics CDMO, March to September 2026. The client is anonymised.