GEO for SaaS: how to get cited by AI answers

Generative engine optimization for SaaS is the work of making your product the one an AI assistant names, and your page the one it cites, when a buyer asks a question like 'what is the best invoicing tool for a two-person agency'. In the 2023 study that named the field, adding quotations, statistics and cited sources to a page raised its visibility inside generated answers by roughly 30 to 40% over an unoptimized baseline, while keyword stuffing did not help [1]. Google says there are no special requirements to appear in AI Overviews beyond being indexed and eligible for a snippet [2]. The practical consequence is that GEO is mostly about clear, checkable content in the places assistants read, and many of those places are other people's sites, which is where partners come in.

What is generative engine optimization for SaaS?

Classic SEO tries to win a position on a results page. Generative engine optimization tries to win a sentence inside an answer. An assistant such as ChatGPT search, Perplexity, Google AI Overviews or AI Mode retrieves a set of pages, reads them, and writes one response that names products and links a handful of sources. Your goal is to be in the retrieved set, to be the passage the model actually uses, and to be credited with a link.

For a SaaS company the questions that matter are narrow and commercial: 'X alternatives', 'best tool for Y', 'X vs Z', 'how much does X cost', 'does X integrate with Stripe'. These questions have short answers with numbers in them. That is good news, because short answers with numbers are exactly what generated responses are built from. It is also a warning: if your pricing page hides the number behind a sales call, the assistant will take the number from somebody else's page, which may be out of date.

The name comes from a paper by Pranjal Aggarwal and five co-authors, first posted in November 2023 and accepted at KDD 2024 [1]. They built a benchmark of queries, rewrote source pages with different strategies, and measured how much of each page made it into the generated answers. We will use their results below, with the caveat that the experiments were run on research setups and one deployed engine, not on every assistant you care about today.

How do AI assistants pick which sources to cite?

Every assistant with web access does two things in sequence. First it retrieves: it runs one or more searches, or reads an index it built with its own crawler, and collects candidate pages. Then it generates: it reads the candidates, writes the answer, and attaches citations to the passages it drew from. You can only be cited if you survive the first step, and you are only cited if you are useful in the second.

Retrieval is where crawler access matters. OpenAI documents three agents: OAI-SearchBot, which surfaces sites in ChatGPT's search features; GPTBot, which collects content that may be used to train models; and ChatGPT-User, which fetches pages when a user asks ChatGPT something. OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, and that blocking GPTBot does not affect search visibility [3]. Google's position is simpler: to be eligible for AI Overviews and AI Mode, a page must be indexed and eligible for a snippet, and the usual controls (nosnippet, data-nosnippet, max-snippet, noindex) limit what appears [2].

The common mistake we see is a blanket 'block all AI bots' rule copied from a blog post, which also blocks the search agents that send traffic. If you want to keep your content out of model training but stay visible in answers, block the training agents and allow the search agents. Check your robots.txt and your CDN's bot settings separately, because many CDNs now have their own AI-crawler toggles that sit in front of robots.txt.

Generation is where content matters. A model asked 'how much does X cost' looks for a passage that states the price. A model asked 'X vs Z' looks for passages that compare the two on named dimensions. A passage that answers the sub-question in one or two sentences, with a number and a date, is easy to lift. A passage that needs three paragraphs of context before it says anything is not.

What makes a page quotable in AI answers?

The GEO study tested nine rewriting strategies. The strongest were adding quotations from relevant sources, adding statistics, and citing sources inline. The authors report that these reached 30 to 40% relative improvement on their main visibility measure (position-adjusted word count), with quotation addition best at about 41%. Fluency and readability rewrites gave a smaller gain of 15 to 30%. Keyword stuffing gave little or no improvement, and on the deployed engine it did about 10% worse than baseline [1].

Translate that into page rules. Put the answer first. If the title is a question, the first paragraph answers it in three to five sentences that make sense on their own, because that paragraph is the one most likely to be lifted whole. Use real numbers. A price, a limit, a percentage, a number of days. Name the source for any number you did not produce yourself and link it. Use headings phrased the way people ask, so the retrieval step can match a sub-question to a section.

A worked rewrite

Here is a typical pricing FAQ answer before: 'Our flexible plans grow with you, and our team is happy to find the right fit for your business. Reach out to learn more about volume options.' Nothing in it can be quoted. Here is the same answer after: 'The Team plan costs $49 a month for up to 5 users; each additional user is $8 a month. Annual billing is $490 a year, which is about 17% less than paying monthly. Prices are in US dollars and were last changed on 1 September 2026.' Every sentence is a fact a model can cite, and the date tells it the fact is current.

Check the arithmetic in your own copy. In the example, twelve months at $49 is $588, and $490 is $98 less, which is 16.7% of $588, so about 17%. Assistants increasingly compare numbers across sources, and an arithmetic error on your own page is a reason for the model to prefer a different source.

Does structured data help you get cited by AI answers?

Less than people hope. Google's guidance for AI features says there is no special schema.org structured data you need to add, and that you do not need new machine-readable files, AI text files or markup to appear in AI Overviews or AI Mode. What it does ask is that structured data matches the visible text on the page [2]. Structured data remains useful for ordinary search features, and it can help any system that parses it to read your prices or product names correctly, but it is not a switch that makes an assistant cite you.

The useful rule is consistency. If your Product markup says $49 and your visible pricing table says $59 because someone updated one and not the other, you have created two conflicting facts about yourself. Pick a single source of truth for prices and terms, render both the visible page and the markup from it, and you will never ship that conflict.

Should a SaaS site publish llms.txt?

llms.txt is a proposal by Jeremy Howard, first published on 3 September 2024, for a markdown file at /llms.txt that gives language models a curated map of a site [4]. The format is small. An H1 with the site or project name is the only required part. It is followed by a blockquote summary, optional detail paragraphs, and H2 sections that list links in the form [name](url) with an optional note. A section titled Optional holds links an agent can skip. The proposal also suggests offering clean markdown versions of pages at the same URL with .md appended [4].

Be clear about what it is and is not. It is a proposal, not a standard adopted by the major assistants, and Google's AI features documentation says you do not need AI text files to appear [2]. Where it does help is with agents and developer tools that are told to read a site: a coding assistant pointed at your API docs, or an agent comparing affiliate programs, gets a short, ordered list of the pages that matter instead of crawling navigation menus. RelayWonder publishes an llms.txt for that reason, and we treat it as a cheap courtesy, not a ranking factor.

• Keep it short: the pages a buyer or an agent needs, such as pricing, plans, integrations, security, terms and docs.
• Write one plain sentence of description per link, with numbers where they exist ("Pricing: three plans from $19 to $199 a month").
• Generate it from the same source as your sitemap so it never lists a page that no longer exists.
• Do not put anything in it that is not also on the visible site. It should be a map, not a second, secret version of your claims.

Why partner content on third-party sites matters more than your blog

Think about the questions that lead to a purchase. 'Best tool for Y', 'X alternatives', 'is X worth it'. An assistant answering these is trying to give a balanced view, and a balanced view is assembled from several independent sources. Your own site is one voice, and on a 'best tool' question it is the least credible voice, because every vendor says it is the best. The citations that carry those answers tend to be review articles, comparison pages, tutorials, community threads and videos by people who use the product.

That makes partner content a GEO asset in its own right. A tutorial by a YouTube creator, a comparison post by a niche blogger, a newsletter issue that walks through a setup, an honest answer in a Reddit thread: each is a separate page that a retrieval step can find and that a model can cite as independent evidence. Each one can also carry a tracked link or a coupon code, so you can see whether the citations turn into customers.

Here is an illustrative way to think about the share of an answer. Suppose an assistant typically cites five sources for a commercial question in your category. That number is an assumption for the example, not a measured average. Your own site can realistically occupy one of those slots. If three independent partner pages also mention you favorably, you appear in up to four of five sources; if none do, your best case is one of five, and the other four slots go to pages that may recommend a competitor.

Two cautions. First, the partner content has to be genuinely useful, because the same quality signals apply to it: direct answers, real numbers, cited sources. A thin page stuffed with your brand name will not be cited and may hurt the partner. Second, paid relationships must be disclosed. The FTC's guidance says an affiliate relationship should be disclosed clearly and close to the recommendation, and that the words 'affiliate link' alone may not be understood by readers [5]. A disclosed review is still a citable review; an undisclosed one is a legal risk for you and the partner.

How do you measure whether AI assistants cite you?

There is no Search Console for AI answers that covers every assistant, so measure it the way you would measure anything new: with a fixed panel of questions, run on a schedule, recorded the same way each time.

1. Write 20 questions a buyer in your category would ask, in their words. Include price, comparison, alternative and integration questions.
2. Pick the assistants your buyers use. Three is a reasonable start.
3. Once a week, run each question in each assistant from a logged-out or clean session, and record whether you are named, whether you are cited with a link, and which other sources are cited.
4. Compute a mention rate and a citation rate, and keep the list of third-party sources that keep appearing. Those are the pages and people you want as partners.

Answers vary from run to run, so do not react to one week. Look at the four-week average and at the list of recurring third-party sources. If the same comparison site is cited in a third of your answers and does not mention you, that is the single most valuable outreach target you have. RelayWonder's InsightWonder integration turns this kind of AI-answer monitoring into a list of content opportunities, and Hunter can then find the people behind the recurring sources.

A GEO checklist for SaaS pages

1. robots.txt and CDN settings allow OAI-SearchBot and Googlebot; training agents are blocked only if you have decided to.
2. Every important page is indexed and not marked nosnippet.
3. The first paragraph of each page answers its title question in 3 to 5 self-contained sentences with at least one number.
4. Prices, limits and terms are in plain HTML text, not only in images or behind a contact form.
5. Each external claim links a primary source; each number has a date or a "last updated" line.
6. Headings are phrased as the questions buyers ask.
7. Structured data, if used, matches the visible page and is generated from the same source of truth.
8. An llms.txt lists the 10 to 20 pages that matter, with one-line factual descriptions.
9. Comparison and alternative pages are honest about where a competitor is better; one-sided pages are rarely cited for balanced questions.
10. At least a few partners have published useful, disclosed content about you on their own sites, and each piece carries a tracked link or coupon.

What to do this week

Spend one day on access and one day on content. On day one, read your robots.txt and CDN bot settings and confirm the search agents are allowed. Pick your five highest-intent pages, usually pricing, the main product page, two comparison pages and your integrations page, and rewrite each first paragraph so it answers the page question with a number. On day two, write the 20-question panel and run it once to get a baseline.

Then look at the third-party sources in that baseline. Make a short list of the people who wrote them: bloggers, creators, newsletter authors, community regulars. Those are your first partner conversations. Generative engine optimization for SaaS is not only about your own pages; it is about being described accurately on the pages assistants already trust.