The ChatGPT Traffic Came Back. What Is It Actually Worth?
A few months ago, an ecommerce client moved to a new hosting setup. Afterwards, ChatGPT citations started falling and AI referral traffic followed. Google rankings held fairly steady, yet organic revenue kept declining.
The usual SEO checks didn't explain it. The server logs did.
The firewall was blocking the real OpenAI crawlers
The new firewall was returning 403 errors to OpenAI's crawlers. Some of those requests came from bots impersonating GPTBot, and blocking them was correct. The problem was that genuine OpenAI requests, verified against OpenAI's published IP ranges, were being blocked as well. In the log sample I reviewed, there were 21 blocked requests to robots.txt and 29 to the homepage.
A site can load perfectly in a browser and hold its Google positions while another system is shut out entirely. Rankings and page content would never have shown this. The access layer did.
We allowlisted the verified OpenAI ranges and kept the impersonators blocked.
The recovery was measurable, within limits
Citations returned. The store is now the most-cited retailer in its category across the buying questions we track, ahead of several much larger names. That describes our tracked question set rather than every search a customer could run.
AI referrals recovered too: more than 800 ChatGPT referrals in the past 90 days, with visitors landing on product pages and reaching checkout.
Organic and direct traffic also rose after the fix. Some customers may have discovered the store through ChatGPT and come back via Google or a direct visit. That journey is plausible. It isn't proven. Other SEO work, promotions, seasonal demand and customer behaviour all move those channels, and restoring OpenAI access does not by itself explain a Google lift. I can report that these happened together. I can't yet assign a share of that growth to AI.
Eight hundred referrals don't answer the ROI question
A citation ranking doesn't tell a business owner what AI search is worth, and neither does a referral count. Checkout activity shows intent, but it isn't a completed purchase.
The next step is following identifiable ChatGPT visits through to orders, then looking at what those orders contribute after the cost of supplying them. The reporting window matters as well: a 90-day referral total shouldn't be compared casually with one month's sales or campaign spend.
Some of the journey will stay invisible to analytics. A customer can ask ChatGPT, check the business on Google, compare products and return days later, and only the last visit gets credited. Asking customers how they found the store, allowing more than one answer and storing it against the order record, fills part of that gap.
That leaves three sets of evidence to keep separate: purchases tied to identifiable AI referrals, purchases where customers report AI influence, and the wider organic and direct changes. De-duplicate the first two. Treat the third as context until there's evidence linking it.
Costs need the same discipline. When the same work improves both Google and AI visibility, forcing every hour into one channel produces a precise-looking report built on weak assumptions. The combined commercial result matters as much as any allocation to AI.
Which identifiable visits became purchases, what those orders contributed, and whether customers say AI helped them choose: those are the questions for the next review.
Read the full breakdown → https://substack.com/profile/3....40709315-tanya-b/not
