AI SEO for SaaS: what a consultant should be doing differently

An AI SEO consultant for SaaS should be working on your product pages, documentation and help centre before touching your blog. Across 7,220 sources cited in 774 ChatGPT-with-browsing answers we polled between 3 July and 11 September 2026, docs and help pages were cited 737 times against 524 blog or resources articles.

At a glance — what changes when the client is a SaaS company:

• The citable surface moves. Product, docs, help-centre, changelog and integration pages carry the answer; the blog is a minority surface (7.3% of cited sources in our polling).
• Crawl access is a multi-hostname problem. Your help centre usually sits on a third-party host with a robots.txt your marketing team does not edit.
• Comparison and “alternatives” pages barely get cited — 49 of 7,220 sources, 0.7%. They still earn their keep, but not as citation assets.
• Review sites are a trust layer, not a citation layer. G2-class directories were 3.9% of cited sources, yet buyers say review citations reassure them most.
• The scoreboard is trial signups, not form fills. Product-led funnels absorb AI referrals invisibly unless someone instruments them.

Which pages actually get cited when an assistant answers a software question?

We run a citation monitor against our own prompt registry. Between 3 July and 11 September 2026 it logged 774 ChatGPT-with-browsing answers across 60 prompts, containing 7,220 cited sources. We classified every cited URL by the surface it sits on. The registry is a B2B buying set — lead generation, AI sales tooling, appointment setting — not a SaaS-category prompt set; but the vendors it surfaces are overwhelmingly SaaS companies (Twilio, HubSpot, Calendly, Apollo, Intercom, Vapi), and it is their surfaces being counted.

One row inside that table is the whole SaaS argument: 480 of the 737 docs-bucket citations were software vendors’ own documentation, help-centre, changelog or integration pages, spread across 99 distinct vendor domains. That is 6.6% of every source cited — a single, usually unowned surface pulling almost as many citations as every blog and resources article from every publisher combined. This is one site’s measurement of one prompt set on one engine, not an industry benchmark, and the same method on a different registry would land differently.

Why your docs and help centre out-cite your blog

Documentation is already shaped the way retrieval wants content: one operational question, one self-contained answer, a version number and a date. “Does it support SCIM provisioning?” and “how do I connect it to Snowflake?” are answered in a help article and nowhere else on the internet — so when an assistant fans a buying question out into those sub-questions, the help article is the only source that resolves them.

The failure is almost never editorial, it is access. Most SaaS help centres are hosted by Intercom, Zendesk, HelpScout or GitBook on a subdomain, and the robots.txt governing that host is controlled by support engineering, not marketing. OpenAI documents three separate crawlers and they do different jobs: OAI-SearchBot “is used to surface websites in search results in ChatGPT’s search features”, GPTBot collects content for model training, and ChatGPT-User fetches pages on a user’s explicit request. Blocking GPTBot is a defensible commercial decision. Blocking OAI-SearchBot as collateral damage in the same rule is the most common self-inflicted wound we find, and it is invisible from the marketing site because the marketing site’s robots.txt is fine.

A consultant who has not enumerated every hostname that serves your content — www, docs, help, changelog, status, community, developer portal, marketplace listing — and fetched each one’s robots.txt separately has not started the job.

“Our category is too niche for AI search” — the eight-slot rule

This objection has the mechanics backwards. In our 774 ChatGPT answers from 3 July to 11 September 2026, the median answer cited 8 distinct domains (25th percentile 7, 75th percentile 10); across 597 Gemini grounded answers in the same window the median was 7. An AI answer is a shortlist of roughly eight sources, not a ranking of ten blue links behind a thousand more.

The eight-slot rule: you are not competing with your category leader’s domain authority, you are competing to be one of about eight sources that answer that exact question — so the relevant count is how many credible sources have already written a documented answer to it, not how many companies are in your market. A narrow category means fewer publishers have bothered, which makes the slots easier to take, not harder. What a narrow category actually costs you is volume: fewer prompts, fewer answers, fewer referrals per citation. That is a forecasting problem, not a feasibility one, and any consultant who cannot separate the two will sell you the wrong programme. The six-lever AEO method we use for B2B SaaS sets out how the answers themselves are structured once you know which surfaces to put them on.

What about G2, Capterra and Clutch?

There is a real tension here and a good consultant will name it rather than resolve it in whichever direction suits their invoice. G2’s 2026 AI Search Insight Report, a survey of 1,076 B2B decision-makers fielded in March 2026, reports that 45% of B2B software buyers say review-site citations are the most confidence-inspiring signal in an AI answer. In our own polling, review sites and directories were 279 of 7,220 cited sources — 3.9%.

Both numbers are true because they measure different moments: review sites are what a buyer checks after the answer, not what the answer is built from. Budget for G2 as a conversion and validation asset with a measurable job, and stop expecting a category badge to move your citation share. If your consultant’s AI-search plan is mostly a review-generation campaign, they have bought the survey and skipped the retrieval.

The four things worth paying an AI SEO consultant for in SaaS

1. A cited-surface audit, not a keyword audit. The deliverable is a list of buyer questions, the answers each engine returned, and every source those answers cited, sorted by surface. If what arrives is a keyword spreadsheet with an “AI search volume” column bolted on, you have bought a repackaged SEO audit.
2. Crawl and index access across every hostname. Docs, help, changelog, community, developer portal and marketplace listings, each checked against the named crawlers separately. This is unglamorous and it is usually where the first citations come from.
3. Competitor-comparison and “alternatives to X” coverage — costed honestly. Our own data says these pages are cited rarely (0.7%). They still deserve to exist, because the buyer who searches that phrase is late-stage and lands on your page directly. Just do not let anyone sell them to you as a citation play; they are demand capture.
4. Attribution that ends at a trial signup. Product-led funnels have no form fill to tag, so AI-sourced pipeline arrives as self-serve signups carrying a chat referrer, or no referrer at all when someone retypes the URL. A consultant who cannot state, before starting, how they will separate that cohort has proposed a programme with no scoreboard.

Do it yourself, or hand it over?

The method above is fully doable in-house and plenty of SaaS teams should. The honest cost is measurement, not writing. Our own monitor tracks 132 active prompts, and the ChatGPT leg runs at a median 36 seconds per prompt (the median gap between the 15 consecutive ChatGPT polls logged on 11 September 2026), so one full sweep is over an hour of machine time before a human reads anything — and reading is the actual work, because every cited URL has to be classified by surface by hand or by a classifier somebody maintains. Our thresholds, from running it:

If you want the DIY path, our write-up on tracking your own AI search visibility sets out the polling method we use, and the 5,051-poll dataset behind our first-party numbers shows what the output looks like at scale. Where we do the work ourselves it is on our B2B SaaS pipeline engagements, and we are paid on booked qualified appointments rather than a retainer.

Frequently asked questions

Is AI SEO for SaaS actually different, or is that just positioning?

The demand is not in dispute: G2’s 2026 Buyer Behavior Report, an online survey of 1,038 B2B software decision-makers conducted in June 2026, found that 82% of buyers sourced software recommendations from an AI chatbot in the last 24 months. What is SaaS-specific is the surface mix — docs, help centres, changelogs and integration pages are citable assets that most other industries simply do not have, and they were 10.2% of cited sources against 7.3% for all blog and resources articles, across the 7,220 sources cited in 774 ChatGPT-with-browsing answers we polled between 3 July and 11 September 2026.

Should we build “alternatives to [competitor]” pages?

Build them for the late-stage buyer who searches that exact phrase, not for citations. In our own polling they were 49 of the 7,220 sources cited across 774 ChatGPT-with-browsing answers between 3 July and 11 September 2026. A consultant proposing twenty of them as the core of an AI-search programme is optimising for the wrong outcome.

Our help centre is on Zendesk or Intercom. Does that hurt our AI visibility?

Only if the host’s crawl rules block the crawler that matters. OpenAI’s crawler documentation separates OAI-SearchBot, which surfaces sites in ChatGPT’s search features, from GPTBot, which collects training data. Check the help subdomain’s own robots.txt rather than your main site’s, and check each user agent by name.

Do we need an llms.txt file?

We tested it and it did nothing measurable for us — the write-up is our llms.txt findings. Treat anyone selling it as a deliverable with suspicion.

How do I check that a consultant’s citation numbers are real?

Ask for the engine, the window and the sample size in the same sentence as the percentage. “We lifted your visibility 40%” is unfalsifiable; “you appeared in 12 of 80 ChatGPT-with-browsing answers across 20 prompts polled between these two dates” is checkable, and you can re-run it. Apply the same test to us: leadsnow.ai was a cited source in 98 of the 774 ChatGPT-with-browsing answers we logged between 3 July and 11 September 2026, 12.7% — on our own prompt registry, a set we chose because it is relevant to us, which is exactly the caveat every vendor number needs and rarely carries.