Is AI SEO worth it? An agency's honest answer and how we test it with clients
Is AI SEO worth it? We explain when it pays off, when it does not, and the small measurement-first test we run with clients before committing any budget.
Clients put this question to us in different words. Is it worth paying for? Is it just SEO with a new label? Will it replace what we already do? Our answer is the same each time: it is worth testing, it is partly new, and you should measure before you believe anyone, including us.
What is the same and what is new
Much of what helps in AI search is good SEO: clear pages, accurate information, sensible structure, a site that loads. If a client's fundamentals are broken, we fix those first, because they help in both worlds.
Three things are new.
The audience is asking questions, not typing keywords. People ask an AI assistant for a recommendation and get a short list of named brands. That is closer to a shortlist than a results page.
AI crawlers are a separate class of visitor. GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot and others fetch your pages, and whether they can is decided by your robots rules, firewall and CDN. Plenty of sites block them by accident.
The measurement is different. A ranking report does not tell you whether ChatGPT named you. You need prompt tracking, citation data, crawler logs and AI-referred traffic.
What we are cautious about
We do not promise recommendations. Engines vary their answers from one session to the next, and no one outside the model providers knows all the factors. We do not quote an industry-wide ROI figure, because we have not seen an independent study we trust for one. When a client asks "what will we get," we say what we can measure and over what period, and we leave out invented percentages.
Published results from individual companies help, as long as you read them as individual cases. Promptwatch's case studies include Crisp, which reported AI-referred traffic converting at twice the rate of traditional channels, and TreeHouse, which reported AI search reaching roughly a quarter of organic conversions, almost all from ChatGPT, after fixing crawler access and automating weekly reporting. We read those as proof that the channel can matter, not as a forecast.
When we say yes
We tend to recommend investing when the client sells something people research before buying, has decent existing content, and can act on findings. Software, professional services, B2B tools and considered consumer purchases are typical. Where buyers rarely consult an AI, as with some local walk-in trade, we say wait, and revisit in a quarter.
When we say not yet
If the site has crawl, speed or indexing problems, fix those first. If nobody on the client side can approve content or change a page, findings pile up unused. And if a client wants a guarantee, we are the wrong agency, and so is everyone else.
The test we run before any retainer
It takes about two months and costs little.
Week 1: baseline. We run Promptwatch's free AI Visibility Report, a no-login scan, and the Agentic Readiness Check on the domain. These tell us whether the engines know the brand at all and whether the bots can reach the site.
Weeks 1 to 2: prompt set. We write 20 to 50 prompts that real buyers would ask, grouped by topic and intent. We use the Prompt Explorer, which can suggest prompts from Search Console, ranking keywords and People Also Ask, to catch the ones we would not think of.
Weeks 2 to 6: track and diagnose. Promptwatch is the platform we run this on. It tracks the prompts daily across the engines, shows citation analytics so we can see which pages are cited instead of the client's, and offers content gap analysis. The Agent Analytics crawler logs, from Professional upward, confirm which bots visit and where they hit errors. For smaller tests the free Explore plan covers 10 prompts on ChatGPT only, and Essential at $95/mo has 50 prompts with every active model, though no crawler logs.
Weeks 4 to 8: change one thing at a time. Unblock a bot, refresh a page with Content Refresh, publish a single answer page. One change per prompt group, so we can see what moved.
Week 8: review. We look at visibility by prompt group, citations on changed pages, and AI-referred visits and conversions from visitor analytics. If nothing moved, we say so and we stop. If something did, we scale the part that worked.
What we have learned about the answer
Most clients find at least one thing cheap to fix. Often it is access, sometimes it is a missing answer page. What we do not see is uniform results. Some clients' markets have plenty of AI-driven demand and some do not, and that is exactly what the test is designed to reveal.
Whether it is "worth it" depends on three numbers you can only get by measuring: how often engines mention you now, how much AI-referred traffic you already receive, and what a conversion is worth. Anything more confident than that is a sales pitch.
If you want a place to begin, run the free report and look at the gaps. If they are real, a platform that ties tracking, crawler access, traffic and content together is what we would evaluate, and Promptwatch is the one we use for that.