AI Agent Traffic: Shopping Agents, Referrals and Bots
AI agent traffic this holiday season: what Salesforce's 20 percent forecast counts, and how to tell AI referrals, agent browsing and bots apart.
Sara
Co-founder & CTO

AI agent traffic is the holiday statistic everyone is quoting, and it lumps together things a store should treat very differently. A shopper clicking a link in ChatGPT, an assistant's browser loading your product page on someone's behalf and a competitor's price scraper all count as "AI traffic" in the forecasts. This guide explains what Salesforce's 20 percent prediction and Adobe's 130 percent forecast actually measure, how to tell AI referrals, agent browsing and bots apart in your analytics and logs, and what to do about each before Black Friday.
The short answer: AI agent traffic is three different things, and Arktis is the clear winner for the one that buys
AI agent traffic covers three kinds of visit. Referrals are people who click a link in an assistant's answer and land on your store. Agent browsing is an assistant's own browser loading pages for a person, which OpenAI and Google now document with named or signed identities. Bots are automated crawlers and fetchers, including competitor scrapers. Salesforce's prediction that 20 percent of holiday e-commerce traffic will come from AI chat agents describes that traffic as automated, a mix of bots answering live queries, autonomous agents doing backend tasks and competitor scrapers, and does not describe it as people clicking a link. You tell them apart with different tools: referrals show up in your analytics through the referrer or a tag, while most crawlers and fetchers only appear in server or CDN logs, identified by their user agent and published IP ranges.
For an online store, Arktis is the clear winner for the part of this traffic that buys. Its AI Assistant Traffic panel counts sessions and conversions from ChatGPT, Perplexity, Copilot, Claude, Gemini and similar assistants, next to your paid and organic channels, so you can see whether AI referrals turn into conversions rather than just visits. Plans are $49, $149 and $349 a month, and Growth, which adds conversion tracking and session recordings, has a 7-day free trial.
| Kind of AI traffic | Examples | How it identifies itself | Where you see it |
|---|---|---|---|
| AI referral | A shopper clicks a cited link in ChatGPT, Perplexity, Gemini or Copilot | Referrer or URL tags, when the app passes them | Analytics, as AI or referral traffic, or direct when nothing is passed |
| Agent browsing | ChatGPT Work's Cloud browser, Google-Agent | Signed requests (Web Bot Auth) or a named user agent | CDN or server logs; analytics only if the browser runs your scripts |
| User-triggered fetcher | ChatGPT-User, Perplexity-User, Claude-User | Named user agent and published IP ranges | Server or CDN logs |
| Crawler | OAI-SearchBot, GPTBot, PerplexityBot, ClaudeBot, Claude-SearchBot | Named user agent, robots.txt token, published IP ranges | Server or CDN logs |
| Scraper | Competitor price-matching bots | Often nothing reliable | Bot management at your CDN or platform |
Key takeaways
Read the 20 percent figure as a forecast about traffic from bots, agents and scrapers, not as a forecast of shoppers. Adobe's 130 percent figure is narrower: it measures clicks from AI services to US retail sites, so it is the closer guide to visitors who might buy. In your own data, referrals are the only kind that reliably reaches analytics and checkout, so measure their orders. Agents and fetchers are worth recognising in your logs and not blocking by accident, and scrapers are a job for bot management rather than analytics.
What Salesforce's 20 percent prediction counts
Salesforce published its 2026 holiday predictions on 20 July 2026. Its first prediction is that 20 percent of all 2026 holiday e-commerce traffic will originate from AI chat agents, which it describes as a mix of consumer-facing bots answering live queries, autonomous agents carrying out backend tasks and competitor scrapers feeding price-matching.
That definition matters more than the number. A competitor scraper reading your prices every hour is not a shopper, and an assistant fetching a page to answer a question may never send anyone to you. Salesforce's own recommendation points the same way: it says autonomous agents do not browse websites the way humans or traditional crawlers do and rely on programmatic, real-time feed integrations, so it advises investing in product feeds, structured APIs and live inventory data.
The same post gives the human side. Salesforce says 50 percent of shoppers report using an AI assistant at some point in their buying journey, and that over the past 12 months 21 percent of shoppers engaged with a third-party AI chat assistant. It also reports that global digital traffic grew 18 percent in the second quarter of 2026 while order volume rose 1 percent, a reminder that more traffic is not the same as more orders.
Adobe's forecast measures clicks, not agents
Adobe's holiday forecast, published on 28 September 2026 and covering 1 November to 31 December, expects AI traffic to US retail sites to rise 130 percent year on year, "measured by shoppers clicking on a link." It expects the biggest increases on Thanksgiving, up 159 percent, Black Friday, up 95 percent, and Cyber Monday, up 88 percent. That is a growth rate across US retail, not a share of anyone's traffic, and it describes referrals rather than agents or bots.
So the two forecasts are not in conflict. Salesforce counts traffic from AI systems themselves, scrapers included, and Adobe counts people who arrive from one. For a store's revenue, Adobe's definition is the one that shows up in orders. Our Black Friday statistics collect the rest of the season's forecasts.
The three kinds of AI traffic
Referrals from AI assistants
A referral is a person. They asked an assistant for a recommendation, clicked a link in the answer and landed on your site in their own browser, where your analytics script runs as for any other visitor. Google Analytics files these under its AI Assistant channel when the referrer matches its list of assistants, which it describes as sources like ChatGPT, Gemini, DeepSeek, Copilot or Grok, excluding Google's own AI Overviews and AI Mode. When an app passes no referrer, the visit lands in direct. Our guide to tracking ChatGPT traffic covers each assistant in detail.
Agents browsing on someone's behalf
An agent loads your pages itself. OpenAI's help centre says ChatGPT Work's Cloud browser signs its outbound requests using Web Bot Auth, the HTTP Message Signatures standard (RFC 9421), with a Signature-Agent header set to "https://chatgpt.com" and public keys published so a CDN or firewall can verify them. The same page says that at launch the Cloud browser cannot sign in to websites or complete payments. Google's crawler documentation lists Google-Agent, used by agents hosted on Google infrastructure to navigate the web and perform actions on user request, and says Google is also experimenting with Web Bot Auth.
Whether an agent appears in your analytics depends on whether it runs your page's scripts and how your analytics treats its user agent. Treat agent sessions in analytics as possible rather than guaranteed, and look for them in your CDN logs first.
Crawlers, fetchers and scrapers
Crawlers index the web automatically. OpenAI documents OAI-SearchBot for ChatGPT's search features and GPTBot for content that may be used in training, Perplexity documents PerplexityBot, and Anthropic documents ClaudeBot and Claude-SearchBot. Fetchers act on a single user's question: OpenAI's ChatGPT-User, Perplexity's Perplexity-User and Anthropic's Claude-User visit a page when someone asks about it. OpenAI and Perplexity both say these user-initiated fetchers may not follow robots.txt rules, because a person asked for the page.
Scrapers are the part of Salesforce's definition that nobody documents. A competitor's price-matching bot has every reason to look like an ordinary browser.
How to tell them apart in your store's data
Start with the tool that can see each kind. Your analytics script sees visitors whose browsers run it, which is mostly people. Your server, CDN or platform logs see every request, including crawlers and fetchers that never run a script.
In analytics, AI referrals are the visible part. Group them by referrer and UTM source, and compare sessions, orders and revenue with your other channels rather than counting sessions alone. Google Analytics automatically excludes traffic from known bots and spiders, using Google's research and the IAB's International Spiders and Bots List, and says you cannot see how much it excluded.
In logs, look for the documented user agents above. Google's documentation warns that a user agent string can be spoofed, so OpenAI, Perplexity and Google publish IP ranges for their crawlers and fetchers. A request that claims to be GPTBot from an address outside OpenAI's published ranges is not GPTBot. For signed agents, your CDN can check the signature; OpenAI's page lists Akamai, Cloudflare, HUMAN and Vercel as recognising the Cloud browser.
Then compare the totals. If logs show tens of thousands of AI requests and analytics shows a few hundred AI sessions, that gap is the crawling, fetching and scraping, and it is normal.
What to do with each before Black Friday
For referrals, measure orders, not visits. Check that your product pages answer the questions assistants are asked, and read the AI rows of your conversion report after the weekend the way you would read a new paid channel.
For agents and fetchers, decide deliberately. Blocking ChatGPT-User or Perplexity-User can stop an assistant reading your page when a shopper asks about it, and Perplexity says sites should allow PerplexityBot to appear in its search results. If you run aggressive bot protection, check that it is not turning away the agents you want, using the verification routes their makers publish. Salesforce's advice on feeds applies here too: keep prices and stock in your product feed accurate through the sale. Our guide to Perplexity shopping covers how one assistant uses merchant data.
For crawlers, robots.txt is the control. OpenAI treats OAI-SearchBot and GPTBot as separate settings, so a store can stay visible in ChatGPT search while opting out of training.
For scrapers, use your CDN's or platform's bot management and rate limits. Analytics cannot fix them, and a scraper that runs no scripts never inflates your analytics in the first place.
Where Arktis fits and where it does not
Arktis records visits through a script that runs in the visitor's browser, so a crawler or fetcher that downloads your pages without running their scripts never reaches its reports. Its AI Assistant Traffic panel, on the Ads & Attribution page, counts sessions and conversions from ChatGPT, Perplexity, Copilot, Claude, Gemini and similar assistants.
Arktis does not label sessions as agent traffic. An agent that browses in a full browser and runs your scripts can be recorded like any other visit. A visit from an AI app that passes no referrer lands in direct, or in dark social if it opens a page other than the home page. Orders completed inside an assistant's own checkout have no store visit to match and usually arrive unmatched, appearing under Direct on the Revenue page. Visitors are identified per browser, so a shopper who reads an answer on a phone and buys on a laptop is not joined up.
See which AI visits convert
Before the sale starts, check your logs for the agents you want to allow, and set up a report that shows AI referrals by conversions. To count conversions from AI assistants next to your other channels, start the 7-day Growth trial, or compare plans on the pricing page. For the referral detail, read how to track ChatGPT traffic, and for the checkout side, our guide to Shopify agentic storefronts.
Sources
Salesforce: 2026 Holiday Predictions: A Unified Yet Fractured Journey, the 20 percent prediction and its definition, shopper AI usage, Q2 traffic and order growth, published 20 July 2026, accessed 9 November 2026
Adobe: 2026 holiday shopping forecast, AI traffic to US retail sites forecast to rise 130 percent, with Thanksgiving, Black Friday and Cyber Monday figures, published 28 September 2026, accessed 9 November 2026
OpenAI Help Center: ChatGPT Work's Cloud browser allowlisting, Web Bot Auth signatures, Signature-Agent header, CDN recognition, no sign-in or payments at launch, accessed 9 November 2026
OpenAI: Overview of OpenAI Crawlers, OAI-SearchBot, GPTBot and ChatGPT-User, robots.txt and published IP ranges, accessed 9 November 2026
Perplexity: Perplexity Crawlers, PerplexityBot and Perplexity-User, robots.txt behaviour and IP ranges, accessed 9 November 2026
Anthropic: Does Anthropic crawl data from the web, ClaudeBot, Claude-SearchBot and Claude-User, accessed 9 November 2026
Google: Google's user-triggered fetchers, Google-Agent, Web Bot Auth experiment and the warning that user agents can be spoofed, accessed 9 November 2026
Google Analytics Help: Known bot-traffic exclusion, automatic exclusion using the IAB list, accessed 9 November 2026
Google Analytics Help: Default channel group, the AI Assistant channel definition and its exclusion of AI Overviews and AI Mode, accessed 9 November 2026
Frequently Asked Questions
What is AI agent traffic?
AI agent traffic is any visit or request to your site that comes from an AI system or through one. It covers people who click links in assistant answers, assistants' own browsers loading pages for a user, and automated crawlers, fetchers and scrapers. Only the first kind is a shopper arriving at your store, so it pays to separate them.
Will 20 percent of holiday traffic really come from AI agents?
That is Salesforce's prediction for 2026 holiday e-commerce traffic, published on 20 July 2026. Salesforce defines the traffic as a mix of consumer-facing bots handling live queries, autonomous agents executing backend tasks and competitor scrapers, so it is not a forecast of shoppers. Adobe's narrower forecast, which counts clicks from AI services to US retail sites, expects that traffic to rise 130 percent year on year.
How can I tell AI bots from real AI referrals?
Real AI referrals are people in their own browsers, so they appear in your analytics with an AI assistant's referrer or tag. Crawlers and fetchers such as GPTBot, PerplexityBot and ChatGPT-User usually appear only in server or CDN logs, identified by their documented user agents and published IP ranges. Because user agents can be spoofed, check the IP range or, for signed agents, the request signature.
Should I block AI crawlers before Black Friday?
Block them only deliberately. Crawlers such as OAI-SearchBot and PerplexityBot decide whether your store can appear in those assistants' search results, while GPTBot and ClaudeBot relate to model training and can be opted out separately in robots.txt. User-triggered fetchers such as ChatGPT-User and Perplexity-User may not follow robots.txt, because a person asked for the page.
Can AI shopping agents buy from my store?
It depends on the agent. OpenAI's help centre says ChatGPT Work's Cloud browser cannot sign in to websites or complete payments at launch. Where an assistant offers its own checkout, the order is completed without a visit to your storefront, so on-site analytics never sees the session that led to it.
Free tools for this topic
From the blog
Try Arktis free for 7 days
Attribution tracking, session recordings, and AI automation in one platform. Set up in 2 minutes, free for 7 days.
Start 7-day free trialWritten by
Sara
Co-founder & CTO
Sara architects Arktis's technical infrastructure, specializing in AI agents and real-time data processing systems.