Somebody on your team has probably already asked the question: should we just block GPTBot and the rest of them. Before you answer that, you need a way to measure AI crawler value on your own site, not a gut feeling and not someone else's case study. This guide walks you through a method for doing that: pulling the AI referral traffic you can already see, building a repeatable AI citation tracking sample, and connecting both back to leads and pipeline, so your answer rests on your own numbers. By the end you will have an evidence pack you can hand to whoever makes the actual block-or-allow call.
What you need: access to your analytics reports, a way to run and record prompts on a few AI engines (by hand or with a tool), and access to your CRM or lead data for the same period.
Get clear on what you are actually measuring
A crawler request is not a visit. A visit is not a citation. And a citation is not a conversion. These three things get talked about as if they are one thing, and that is where a lot of the "should I block AI bots" debate goes sideways.
Here is the chain you are actually trying to measure: access, then visibility, then outcome. Access means the crawler can fetch your pages, which is a technical prerequisite and nothing more. Visibility means your brand gets mentioned or a page of yours gets cited when someone asks an AI engine a question your buyers actually ask. Outcome means a real person showed up and did something you can count: a visit, a lead, a sale.
OpenAI is a good example of why this distinction matters in practice. OAI-SearchBot is the crawler tied to showing up in ChatGPT search. GPTBot is tied to content that may train the underlying models. ChatGPT-User shows up for certain user actions and has nothing to do with whether you appear in search. Blocking one of these does not automatically answer the question for the others. The configuration in your robots.txt file is what actually controls that access, engine by engine, so before you measure anything, get specific about which access, which engine, and which question you are actually trying to answer this quarter.
Done when: you can state, in one sentence, which crawler or engine behavior you are evaluating and why, instead of talking about "AI bots" as one lump category. That single sentence is also what turns a vague sense of AI crawler ROI into something you can actually go measure.
Set your measurement question and reporting window
Pick one question and write it down. Something like: "During this quarter, did AI exposure and AI-referred visits produce meaningful qualified activity compared with our other channels." That is the question this whole exercise answers. Resist the urge to widen it once you start looking at numbers.
Then lock in the details before you look at a single report: the reporting period (match your normal marketing cadence, don't invent a special one), which engines you'll track separately, the buyer prompts and funnel stages you care about, the pages you're watching, which existing conversion events count as outcomes, your CRM's definition of a qualified lead if you have one, the channel you're comparing against (organic search is the usual pick), and who owns reviewing the results.
If your site gets very little AI traffic right now, give it more time rather than lowering the bar. A month of near-zero data is a real answer. A week of near-zero data mostly just means you haven't looked long enough yet.
Common mistake: changing your prompt list, your reporting window, or what counts as a conversion partway through the exercise. Once you do that, you no longer have a before-and-after you can trust, you have two different measurements you're pretending are one.
Pull the AI referral traffic you can already see
You don't need to rebuild your analytics setup for this. Use what you already have and pull, by platform where you can tell them apart: sessions, users, landing page, new versus returning visitors, whatever engagement metric your team already tracks, pages per session, existing lead events, existing conversion events, conversion rate, and the date range.
Keep each platform separate rather than lumping "AI traffic" into one number. Someone arriving from ChatGPT behaves differently than someone arriving from Perplexity or Gemini, and averaging them together hides exactly the differences you're trying to see. This is the AI referral traffic tracking layer of the method, and it's worth doing platform by platform even if the totals look small at first. Without this layer in place, any claim you make about AI crawler ROI is really just a guess dressed up as a number.
OpenAI has said that sites allowing OAI-SearchBot can track referral traffic from ChatGPT in their analytics, and that ChatGPT referral URLs sometimes carry a source parameter that helps identify the visit. Google folds its AI Overviews and AI Mode appearances into its standard Search Console reporting, so that report is worth checking alongside your analytics too. Neither guarantee means every AI-influenced visit keeps its referrer intact. Treat what you see here as a floor, not the whole picture. We'll come back to that in a later step.
Done when: you have a table of visible AI-referred visits broken out by platform, period, landing page, and outcome, using the same definitions you already use for every other channel report. You should be able to answer: which platforms sent visits, which pages they landed on, how many became leads or conversions, and whether the trend is growing, flat, or shrinking.
Where people go wrong: comparing raw AI conversion counts against raw organic conversion counts when the volumes are wildly different sizes. Look at rates and quality, not just totals, or you'll draw the wrong conclusion from a small sample. Good AI referral traffic tracking is what makes that comparison possible in the first place.
Build a repeatable AI citation sample
This is where AI citation tracking comes in, and it's the piece most teams skip because it feels harder to pin down than a traffic report. It doesn't have to be. Write down twenty or so prompts your actual buyers would ask, spread across four kinds of questions: category and educational questions, problem or use-case questions, comparison questions, and product or decision-stage questions. Run each one on the engines you care about, using the same wording, country setting, and language every time.
For every prompt run, record the date, the engine, whether your brand got mentioned, whether your site got cited, the exact URL that got cited (not just the domain), whether a competitor showed up instead, and their cited URL if so. Run each prompt at least three times over five to seven days before you draw any conclusion from it. One answer on one day tells you almost nothing, because AI answers shift by engine, by country, and sometimes by the hour.
A few numbers are worth tracking from this sample. Your mention rate is the share of prompt runs where your brand shows up in the answer at all. Your citation rate is the share of runs where at least one of your pages gets cited, calculated as site citations divided by total prompt runs. Page citation share tells you which specific URLs are doing the work. Calculate share of voice separately per engine rather than blending them, since different engines cite differently. And track competitor share of voice the same way, keeping mentions and citations as separate counts rather than folding them together.
This is one of the two or three places in this method where a tool genuinely saves you the manual work. DeepSmith's AI Visibility module runs your tracked prompts on a schedule and reports mention rate, citation rate, share of voice, and the exact pages being cited, across the AI engines your plan covers. It turns this step from a recurring chore of manual searches into a standing report you can pull up any time. It does not replace your analytics or your CRM, and it will not tell you that a citation caused a conversion. What it gives you is a repeatable record of the visibility layer, which you then join to the traffic and pipeline data from your own systems in the next two steps.
Common mistake: treating one manual ChatGPT search as your citation rate. AI answers vary by engine, country, day, and even how the prompt is worded. A single screenshot is a story you tell yourself. A repeated, recorded prompt set is a measurement.
Join citations to the pages people actually land on
The exact page URL is the thread that ties this whole method together. For every page that showed up cited in your prompt sample, check it against the referral report you built two steps ago. Was that same page also where AI-referred visitors landed? Did visitors land somewhere else instead? Which pages get cited repeatedly but never show up in your referral data? Which pages get AI referrals but never show up as cited in your prompt sample?
Organize this by page, by engine, and by prompt cluster, and you'll start to see four different kinds of value show up. Some pages have citation-only value: they're visible in answers but you can't observe anyone actually landing there. Some have referral value: people are showing up. Some have conversion value: those visitors are triggering something you already track as a conversion. And a smaller set will have pipeline value: those conversions are turning into qualified leads or real deals.
Do not assume a page being cited and a page being landed on are the same event. Someone reading an AI answer might remember your brand, search for you later, and land on your homepage instead of the page that was actually cited. That's still worth counting, just not as a direct hit on that specific URL.
DeepSmith's page-level AI visibility reporting is useful here too, since it hands you the cited-page and prompt context in one place rather than you cross-referencing a spreadsheet against your analytics by hand. Your analytics and CRM stay the source of truth for the actual visits and outcomes; the tool just saves you the manual matching.
Done when: you have a page-level table with columns for the URL, prompt cluster, engine, citation count, AI referral visits, leads, qualified leads, conversions, and a notes column for anything about the attribution that isn't clean.
Connect AI-referred visits to leads and pipeline
Split what you find into two tracks, and do not let them blur into each other. Track A is directly attributed: a lead or conversion your analytics or CRM ties to an identifiable AI-referred visit. For this track, report AI-referred sessions, leads, qualified leads, opportunities if your CRM tracks them, conversions, and the conversion and qualification rates by platform. The two ratios worth calculating are straightforward: AI-referred conversion rate is AI-referred conversions divided by AI-referred sessions, and AI-referred lead rate is AI-referred leads divided by AI-referred sessions.
Track B is everything else: outcomes where AI may have played a part but the final visit wasn't recorded as AI traffic. A "how did you hear about us" answer that names ChatGPT. A CRM field a rep filled in after a call. Sales notes that mention a prospect referencing something an AI tool told them. A bump in direct or branded traffic that lines up with a citation trend you're already tracking. Label these as self-reported, assisted, or directional, and keep them in a separate column from Track A. Never merge the two.
Where people go wrong: assuming a citation caused a conversion just because the cited page later picked up a lead. That's an association at best. Treat it as evidence to note, not a claim to make.
Account for the AI traffic you can't see
Not every AI-influenced visit shows up cleanly in your analytics, and that's worth planning for rather than getting surprised by. Someone reads an answer, remembers your brand, and comes back directly a few days later, showing up as direct traffic with no trace of the AI answer that started it. Ahrefs studied this at scale and found that a meaningful share of AI-driven visits end up looking like direct traffic in analytics, which means whatever you measured in your referral report earlier is a floor, not the full amount.
Give yourself three honest labels for the reporting period: observed direct, where the source is clearly recorded as an AI platform; observed indirect, where a person or your CRM tells you AI played a role but the session itself landed as something else; and unresolved, where you have reason to think AI exposure happened somewhere in the journey but no reliable way to connect it. That third bucket is not a failure of your measurement, it's an honest accounting of what you cannot see yet.
A few checks help here: compare your visible AI-referral trend against your citation trend for the same prompts, look at whether cited pages pick up direct or branded traffic afterward, and compare self-reported AI mentions against what your analytics attributed directly.
Pro tip: a low visible referral count doesn't automatically mean AI traffic isn't worth much to you. It might just mean the platform or the visitor's later path didn't preserve a referrer. Use citation trends and self-reported discovery as supporting evidence, but never write them up as if they were directly attributed conversions. That distinction matters when you eventually make the case for or against blocking access.
Compare quality against your baseline, not just volume
A small AI channel can matter if the people it sends are unusually likely to convert. A large one can matter less than it looks if it never turns into anything. Line AI traffic up against your existing baseline, usually organic search, using the exact same metrics: sessions, your standard engagement measure, pages per session, lead rate, qualified-lead rate, opportunity rate if you track it, conversion rate, and new versus returning visitors. Keep platforms separate here too, since blending them into one AI average can hide a real difference between, say, ChatGPT and Gemini traffic on your own site.
What a useful conclusion sounds like: "Platform A sends fewer visits than organic search, but the ones it sends convert at a higher rate." Or: "Platform B lands on our highest-intent pages, but we don't have enough conversions yet this quarter to say anything definitive." What an unhelpful conclusion sounds like: "AI traffic must be worth it because some other company's report says its conversion rate is higher than organic." Numbers from an industry study are context for reading your own data, not a substitute for it. Semrush's own research found AI-search visitors converting several times higher than typical organic visitors across a large sample of topics, and that is a real, useful data point, but it describes that study's sample, not automatically your site.
Package the evidence and hand it off
The output of all this is not a decision. It's an evidence pack that a separate block-or-allow framework can act on, and building that decision matrix is its own piece of work, not something to improvise at the end of a measurement exercise. Pull together four summaries: visibility (your prompt count, engines, mention rate, citation rate, which pages are earning citations, and competitor share of voice), traffic (visible AI referrals by platform, landing pages, engagement, and directly attributed leads), pipeline (leads, qualified leads, opportunities, conversions, split by directly attributed versus influenced), and confidence (label each finding as strongly observed, observed but limited, influenced or directional, or unknown).
That confidence label matters more than people expect. A strongly observed finding, direct source tied to a direct outcome, carries a different weight than a directional one built from self-reported forms and coinciding trends. Keep both in the pack, just labeled honestly, so whoever makes the actual call on whether to block AI bots on your site is working from a clear picture instead of a blended, over-confident one. That is the whole point of learning to measure AI crawler value in the first place: turning "should I block AI bots" from a debate into a question your own data can answer.
If you want to see the full range of what teaching-first product context and repeatable citation tracking looks like without building the reporting yourself, start a free trial with DeepSmith. It won't make the block-or-allow decision for you, but it will keep the citation half of this evidence pack current without you doing a manual prompt run every few weeks.
