
On this page14
- What Did Google Publish?
- How Does Google Say You Can Verify Google-Agent?
- The Six Ways Allowed Traffic Still Gets Blocked
- What the Document Does Not Say
- Why This Is Not Only a Security Story
- Access Is Half the Picture
- A Two-Week Check for Any Client
- A Critical Look at “Just Allow Everything”
- Which Clients Should Look First
- What a Mention Check Looks Like in Practice
- Why One Check Is Not Enough
- What to Ask Any Vendor
- What to Tell Clients
- Frequently Asked Questions
Google published a help page on letting Google-Agent, its user-triggered browsing agent, through CDNs, firewalls and bot filters. The page lists six reasons allowed requests still get blocked. It covers access only. It says nothing about whether an answer names your client, and that second question is where AI visibility tracking begins.
What Did Google Publish?
Search Engine Roundtable reported on November 17, 2026 that Google released a new help document on how to allow Google-Agent traffic. Google describes the agent as “the name used by agents hosted on Google infrastructure to navigate the web”. It acts when a user asks it to.
Google-Agent is not new. The same article notes that Barry Schwartz covered it in March. What is new is the first guide aimed at site owners who want to let it in. Google’s own page says that explicitly allowing the agent in a CDN, firewall or bot-detection service helps prevent legitimate requests from being blocked, while existing security controls stay in place.
Read that sentence twice. Google is not asking sites to drop their defenses. It is asking them to add one narrow allow rule and to place it correctly.
How Does Google Say You Can Verify Google-Agent?
The page gives three checks. The request carries “Google-Agent” in the user-agent header. The source address sits in a published list of user-triggered agent ranges. And, where present, a Web Bot Auth signature validates. Google describes the signature as an experiment, tied to an identity it is testing.
The Search Engine Roundtable article does not reproduce these details. It shows a screenshot and a summary. So the checks above come from Google’s page itself. Anyone changing a firewall should read that page, not a news recap.
The Six Ways Allowed Traffic Still Gets Blocked
This is the useful part of the document. Most teams that add an allow rule assume the job is done. Google lists six reasons it may not be.
| Reason | What it means in practice |
|---|---|
| Rule order | The allow rule must run before broad cloud or data center blocks, managed bot rules and AI agent block policies |
| Interactive challenges | The rule must also skip CAPTCHAs and JavaScript challenges, including those triggered by search or form use |
| Actions beyond the homepage | Search, filters, detail pages, add-to-cart, booking and the API endpoints behind them all need coverage |
| Shared-IP rate limits | Many users share one egress pool, so per-IP limits can throttle legitimate agents |
| Stale IP ranges | Ranges change, so rules built on old lists drift out of date |
| Proxy headers | Load balancers must pass on the original IP, the user-agent and any signature headers |
Notice how ordinary these are. None is exotic. A site can pass a quick test on the homepage and still fail the moment an agent tries to add an item to a cart. That is the trap.
What the Document Does Not Say
The omissions matter as much as the content.
The page does not mention robots.txt. Search Engine Roundtable also says it does not explain whether Google-Agent follows robots rules. The page does not cover login or account flows. It does not say how often agents visit or what share of traffic they will form. And it does not say that allowing the agent changes how Google ranks or describes a site.
So this guide does not claim any of those things. A reasonable agency can still act on what is stated. It cannot promise a client that an allow rule will win more traffic, more mentions or more sales.
Why This Is Not Only a Security Story
Here is the inference, labeled as one. As agents browse for users, a site that blocks them by accident becomes invisible to those users at the moment of action. A buyer who asks an agent to find a booking slot may never see a client whose firewall turns the agent away. No report quantifies how often that happens. It follows from how the page describes the block points.
That reframes a technical task as a client risk. The firewall belongs to the developer. The risk belongs to the account manager. Clients who ask how to appear in ChatGPT answers should hear this too, because an unreachable site cannot be read by anyone.
Access Is Half the Picture
Allowing an agent through answers one question: can it reach the page? It does not answer the next one: what does it, or any assistant, say about the client when a buyer asks?
AI visibility tracking covers that second half. It asks fixed buyer questions on a schedule, stores each answer with its date and records whether the client was named. It reads answers. It does not open firewalls, and it does not test what an agent does on a checkout page. SKYA states that limit plainly in its page on daily AI answer readings.
Put the two halves together and the full chain appears. An agent must reach the site. An assistant must name the client. The facts in the answer must be right. A weakness in any link breaks the result. Good AI visibility tracking tells you which link failed on a given day.
A Two-Week Check for Any Client
-
Ask the client’s developer or host which CDN, firewall or bot service is in use. Record the names.
-
Open the rule list. Find any rule that blocks cloud ranges, data centers, automated visitors or AI agents. Note its position in the order.
-
Read Google’s page. Compare each of the six points with the client’s setup. Mark each pass, fail or unknown.
-
Check server logs for Google-Agent visits. Note whether they reach product, search and booking pages or stop at the first page.
-
Walk the main buyer journey as a guest. Count CAPTCHAs, forced sign-ins and pop-ups.
-
Choose ten buyer questions. Ask each in at least three assistants. Save each answer with the date and mark whether the client was named.
-
Write the findings and the options. Let the client decide what to allow.
-
Repeat the answer check after two weeks and compare.
Step seven is the one agencies skip. A rule that blocks agents may exist for a reason: cost, security or policy. The agency’s job is to show the trade-off, not to remove the rule.
A Critical Look at “Just Allow Everything”
Some advice online will say to open every door to every agent. This guide does not. Three cautions apply.
First, a broad allow rule can undo protections the client chose on purpose. Google’s own wording says to keep existing security controls in place, and a rule placed above everything else can weaken them if it matches too widely. Second, one allowed agent is not every agent. The page covers Google-Agent. Other assistants use other agents, with their own rules and their own proof of identity. Third, the document gives no evidence that being allowed leads to a gain. It is a guide to avoiding wrongful blocks, not a growth tactic.
The sensible position sits in the middle. Know what the firewall does. Allow what the client wants allowed. Record the decision. Then measure what buyers actually see.
Which Clients Should Look First
Clients with actions that agents might perform come first: bookings, orders, quote requests and appointment searches. Retail, travel, local services and clinics belong here. For them, a block on a cart or booking endpoint costs a real sale.
Publishers and information sites come second. An agent that cannot read an article cannot summarize it. Clients behind heavy login walls come last, since the page does not cover authenticated journeys.
Clients in the EU deserve a separate note. Security and data rules there are often stricter, so the developer and the client’s legal contact should review any change before it goes live. The document does not discuss regional rules.
What a Mention Check Looks Like in Practice
Many agencies say they check assistants and mean they ask one question once. A real check is more structured.
Write ten questions in the words buyers use. Cover each stage: a broad “best option for” question, a comparison, a price question and a “where do I book” question. Ask each on three different days, because answers vary. Save the full text. Then read each answer for four things: whether the client is named, where it sits, how it is described and whether any claim about hours, price or booking is right.
This is also the honest answer to a client who asks how to appear in ChatGPT answers. First learn whether the client appears at all. Then learn why. SKYA’s guide on appearing in ChatGPT answers explains how earned mentions and ads differ. A firewall rule does not change who ChatGPT names, so do not sell it as a fix for that.
Why One Check Is Not Enough
Firewall rules change. Providers update managed rules. IP lists move. A page that passed in November can fail in January after a routine security update. Answers change too.
A series beats a snapshot. Monthly access checks, plus weekly stored answers, show whether a change on the site lines up with a change in what assistants say. That link is not guaranteed. It is worth watching. SKYA’s note on technical AEO and crawlability covers the crawl-side checks, and its guide to the llms.txt file covers a way to give machines a clear summary of a site.
What to Ask Any Vendor
Some vendors sell “agent readiness” scores. Ask what sits behind them.
| Requirement | Why it matters | How SKYA handles it |
|---|---|---|
| Separates access from answers | A site can be open and still go unnamed | Reads answers only and states that it does not test firewalls or checkouts |
| Keeps answer text with dates | A mention without a copy cannot be tested | Every reading is stored with its date, and each figure says when it was read |
| Reads many assistants | Each one names different clients | Nine assistants every day: ChatGPT, Gemini, Perplexity, Claude, Google AI Overview, Google AI Mode, DeepSeek, Grok and Copilot |
| Admits what it cannot read | A guessed score hides gaps | No answers read means no figure. A run that measured nothing gives its credit back |
| Fits agency work | Many clients, one report | Multi-client workspaces, GA4 and Search Console integration and white-label PDF reports |
SKYA publishes how every number is calculated. Ask any vendor to say plainly what it measures and what it leaves out.
What to Tell Clients
Keep it short and factual. Google published a guide on letting its browsing agent through firewalls. We checked your setup against it. Here is what passes, here is what fails and here is what we suggest. You decide what to change.
Do not promise more traffic. Do not warn of lost sales. The honest position is that the guide is new, the details beyond access are thin and a check costs little. Then add that the answer side is measured separately, with dated records, so the client sees both halves in one report.
Frequently Asked Questions
What is Google-Agent?
Google describes it as the name used by agents hosted on Google infrastructure to navigate the web and act on a user’s request. It is not new, but Google has now published a guide to allowing it.
How do I allow Google-Agent on my website?
Add an explicit allow rule in your CDN, firewall or bot-detection service, place it above broad blocks and verify requests using the user-agent, the published IP list and, where present, the signature. Read Google’s page for the exact steps.
Does Google-Agent follow robots.txt?
The help page does not mention robots.txt, and Search Engine Roundtable says the coverage does not explain it either. Check Google’s documentation before assuming either way.
Will allowing Google-Agent help my client appear in AI answers?
Nothing in the coverage says so, and it does not explain how to appear in ChatGPT answers either. The guide is about avoiding wrongful blocks. Whether a client is named in answers is measured separately, using stored readings across assistants.
What does AI visibility tracking add to an access check?
It shows what assistants actually say about the client, with dates and answer text. An access check shows whether an agent can reach the site. Both are needed, and neither replaces the other.
Opening the door is step one. Seeing who gets named is step two. Start a 7-day free trial with SKYA and read the stored answers. No card required.
Written by
SKYA Editorial Team
The SKYA Editorial Team publishes field-tested research on Answer Engine Optimization, Generative Engine Optimization and AI Visibility. Every article is reviewed by senior practitioners from the SKYA product and research groups before publication.
Related posts
Keep exploring the SKYA Journal.
