Written by: Arjun Karnik, Growth Marketing Specialist
Key Takeaways for Agent-Mediated Discovery
- Human search is shifting to agent-mediated discovery, where brands are either cited in AI answers or invisible, with no second page to land on.
- An agent card is a machine-readable JSON file at /.well-known/agent-card.json that lets other AI agents discover, evaluate, and delegate tasks to your brand’s agent without hard-coded integrations.
- Technical plumbing must be fixed first: unblock AI crawlers, add schema markup, publish an Agent Card, and ensure pages are machine-parseable without JavaScript.
- Map fan-out queries from AI prompts, align buyer language in slugs, titles, and headings, then publish structured pages at machine cadence to earn citations.
- Activate a freshness loop to prevent impression decay and run your 30-day A2A pilot with Arjun Karnik and AI Growth Agent.
Agent Card Basics with Copy-Paste JSON
Agent Cards are JSON discovery files that let AI agents locate, evaluate, and delegate tasks to each other without hard-coded integrations. Every A2A-compliant agent publishes its Agent Card at a well-known URL following RFC 8615 so client agents can discover it via unauthenticated HTTP GET requests.
The minimal valid Agent Card JSON below is copy-paste ready. Serve it at https://yourdomain.com/.well-known/agent-card.json with Content-Type: application/json and Access-Control-Allow-Origin: * headers.
{ "name": "Acme Marketing Agent", "description": "Answers questions about Acme's B2B demand generation services and qualifies inbound leads.", "version": "1.0.0", "supportedInterfaces": [ { "url": "https://yourdomain.com/a2a", "protocolBinding": "JSONRPC", "protocolVersion": "1.0" } ], "capabilities": { "streaming": true, "pushNotifications": false }, "defaultInputModes": ["text/plain", "application/json"], "defaultOutputModes": ["text/plain", "application/json"], "skills": [ { "id": "lead-qualification", "name": "Lead Qualification", "description": "Qualifies inbound prospects against ICP criteria and routes to CRM.", "tags": ["b2b", "lead-gen", "crm", "qualification"] } ] }The A2A protocol defines two valid paths for serving an Agent Card: /.well-known/agent-card.json (v0.3.0+, preferred) and /.well-known/agent.json (original, also supported). A2A v1.0 introduced Signed Agent Cards using detached JSON Web Signature (JWS) so clients can cryptographically verify the card issuer, closing impersonation risks that existed in 0.x versions.
Agent Card discovery checklist:
- File is reachable via unauthenticated HTTP GET at
/.well-known/agent-card.json - Response header includes
Content-Type: application/json - CORS header
Access-Control-Allow-Origin: *is set - Card includes at least one skill with
id,name,description, andtags - JWS signature is attached for production deployments
Cache-Control: max-age=3600header is present
See how AI Growth Agent audits and publishes your Agent Card as part of a 30-day A2A pilot.
A2A vs MCP: Picking the Right Protocol Mix
Before implementing your Agent Card, you need a clear view of how A2A and MCP fit together in your stack. Many marketers treat them as competing options, yet most production systems rely on both.
The Model Context Protocol (MCP) focuses on connecting LLMs with data and external resources, while A2A enables agents to collaborate as autonomous entities in multi-turn interactions such as negotiation. MCP operates vertically as agent-to-system for governed tool and API access, while A2A operates horizontally as agent-to-agent for coordination and delegation between multiple specialized agents in multi-agent pipelines and workflows.
Decision checklist for marketers:
- If a second independent agent must share task state and lifecycle, introduce A2A.
- If the workload is one model calling discoverable tools, start with MCP.
- If the workflow spans multiple organizations or cloud boundaries, A2A is required.
- Production stacks typically use both: A2A between agent crews and MCP for agents-as-tools inside a crew.
Step 1: Run the Visibility Audit
Start by establishing a baseline of what AI assistants currently say about the brand. This baseline becomes the control group for every later result.
Run branded prompts (“What is [Brand]?”), category prompts (“Best tools for [use case]?”), and comparison prompts (“How does [Brand] compare to [Competitor]?”) across ChatGPT, Google AI Overviews, Perplexity, and Gemini. For each response, document every mention, absence, competitor recommendation, sentiment signal, and factual error so you can see the full landscape. Record which pages, if any, are cited as sources, because citation tracking starts here. Log the audit date so decay can be measured against it at 30 days. Flag any incorrect AI statements about the brand for immediate defensive GEO correction, because a wrong AI answer hurts more than no answer.
Citation tracking starts here. Share of answer across the four surfaces replaces rank position as the headline KPI. The core measurement framework covers AI visibility (frequency of brand mentions), citation performance (how often the domain is cited as a primary or secondary source), brand representation (accuracy and sentiment), and AI-influenced outcomes (branded search lift and assisted conversions).
Step 2: Fix Technical Plumbing for Agent Retrieval
If AI crawlers cannot read the site, every downstream investment is wasted. The technical plumbing outlined in the key takeaways must be completed before any content strategy.
Technical plumbing checklist:
- Confirm AI crawlers (GPTBot, Google-Extended, PerplexityBot, ClaudeBot) are not blocked in
robots.txt. - Add Organization, Article, FAQPage, and HowTo schema markup across all relevant pages so models can parse context.
- Verify pages are machine-parseable without JavaScript rendering, using server-side HTML for core content.
- Publish the Agent Card at the well-known path defined earlier.
- Submit an updated XML sitemap and implement IndexNow for rapid indexing of new content.
- Add an
llms.txtfile communicating content hierarchy and most important pages to AI crawlers. - Monitor bot-request success rates and 4xx/5xx error rates by crawler type to catch technical failures early.
Step 3: Map Fan-Out Queries and Align Buyer Language
A single buyer prompt triggers dozens of hidden retrieval queries underneath. Focusing only on the visible prompt while ignoring the fan-out targets the wrong surface.
On Arjun’s own site, pages rewritten to match extracted ChatGPT fan-out queries earned citations while control pages did not. That shift came from matching the machine’s internal questions, not just the human phrasing.
Fan-out query mapping workflow:
| Buyer Prompt | Fan-Out Query (Machine Language) | Buyer-Language Relabel | Target URL Slug |
|---|---|---|---|
| How do I get found in AI search? | generative engine optimization techniques | How to get your business recommended by AI search | /how-to-get-recommended-by-ai-search |
| What agent protocol should I use? | A2A vs MCP decision framework | A2A vs MCP: which protocol does your brand need? | /a2a-vs-mcp-decision-framework |
| How do AI agents find my business? | agent card JSON example B2B | Agent card JSON example for B2B marketers | /agent-card-json-example-b2b |
Extract fan-out queries directly from ChatGPT rather than inferring them from keyword tools. The target is the machine’s questions, not the human’s typed query. Align slug, title, H1, and every H2 to the buyer-language column. On Arjun’s own site, relabelling a jargon page to buyer language produced citations within weeks of that specific change.
Step 4: Publish Structured Pages at Machine Cadence
Volume and cadence set the floor for AI visibility. On Arjun’s own site, running AI Growth Agent at 5 to 8 autonomous actions per day, combining new articles with updates to existing ones, took the GEO subfolder from zero to the only source of new impressions on the entire domain in 60 days.
New articles reached thousands of monthly Google impressions within weeks. That output came from consistent, structured publishing rather than a one-time content sprint.
Structured publishing requirements for every page:
- Fan-out query language in the URL slug, title tag, H1, and primary H2s.
- Answer-first formatting, with the direct answer leading and support details following.
- Schema markup on every page (Article, FAQPage, HowTo as appropriate).
- One claim per sentence so a model can lift a single line cleanly.
- Crisp definitions stated plainly within the first 200 words.
See AI Growth Agent run structured publishing at machine cadence on your domain.
Step 5: Set Up the Freshness Loop and Self-Healing Content
Freshness drives ongoing AI visibility rather than acting as simple hygiene. In Arjun’s own tests, pages can drop 78% to 99% in two months without updates, and the position disappears before a monthly report surfaces the problem.
Seer Interactive analyzed 47,097 AI citations across 7,683 pages in ChatGPT, Gemini, and Perplexity between March and June 2026 and found that 75% of cited pages had been updated within the last year, with consistently cited pages averaging under six months since their last update.
Impression-decay tripwire checklist:
- Connect Google Search Console to an automated monitoring layer via AI Growth Agent, which continuously tracks impression trends across all published pages.
- Set a decay threshold (for example, 20% impression drop over 14 days) that triggers an update queue when any page crosses it.
- When a page crosses the threshold, auto-queue the affected URL for a structured refresh with an updated introduction, new data points, and revised H2s aligned to current fan-out queries.
- Immediately after publishing the refresh, re-submit the refreshed URL via IndexNow to accelerate re-indexing.
- Log the refresh date and re-measure citation rate at 30 days to confirm the update restored visibility.
The tripwires fire and updates queue without anyone auditing a spreadsheet. Content repairs itself on a loop instead of waiting for a quarterly audit.
Step 6: Configure Lead-to-CRM Workflow and Audit Trails
Agent-referred traffic converts differently from cold search traffic, so it needs its own workflow. Referred leads convert at 3 to 4 times the rate of cold leads and have about 37% higher retention. The workflow must capture this traffic class separately and route it with appropriate speed.
| Workflow Stage | Action | Metric to Track |
|---|---|---|
| Traffic identification | Segment sessions from chatgpt.com, perplexity.ai, and equivalent AI referrers in GA4 as a distinct traffic class | AI-referred sessions vs. total sessions |
| Lead capture | Route AI-referred visitors to a dedicated landing page with a direct CTA matching the agent answer context | AI-referred conversion rate |
| CRM entry | Tag every lead with source = AI-referral and the originating platform (ChatGPT, Perplexity, etc.) | AI-sourced leads in pipeline |
| Audit trail | Log the citation URL and prompt context that drove the visit where capturable | Citation-to-lead attribution rate |
| Follow-up speed | Trigger immediate follow-up sequence, because AI-referred prospects arrive pre-educated and expect faster engagement | Time from lead to first contact |
Measured impact understates real impact. A meaningful share of AI-driven demand lands in analytics as direct or branded search because buyers copy a name from an AI answer and type it into a browser. Whatever is measured is a floor, not a ceiling.
Step 7: Measure Agent Referrals, Citations, and Share-of-Answer
30-day measurement cadence:
| KPI | Measurement Method | 30-Day Target |
|---|---|---|
| Share of answer | Run 20+ branded and category prompts across ChatGPT, Perplexity, Google AI Overviews, Gemini, and log citation rate | Establish baseline and improve vs. Day 1 audit |
| AI-referred sessions | GA4 segment filtering chatgpt.com, perplexity.ai, and equivalent referrers | Identify and isolate as a distinct channel |
| AI-referred conversion rate | Goal completions from AI-referred sessions divided by total AI-referred sessions | Benchmark against non-AI organic conversion rate |
| Impression decay rate | Google Search Console, week-over-week impression change per URL | Zero pages crossing the decay tripwire threshold without a queued update |
| Agent Card discovery | Server logs, successful GET requests to /.well-known/agent-card.json |
Confirm crawlers are fetching the card |
30-Day Pilot Roadmap
- Days 1–3: Visibility Audit. Run branded, category, and comparison prompts across all four surfaces. Document the baseline. Flag incorrect AI statements for defensive correction.
- Days 4–5: Technical Plumbing. Unblock AI crawlers in robots.txt, add schema markup, publish Agent Card at
/.well-known/agent-card.json, and add llms.txt. - Days 6–8: Fan-Out Query Mapping. Extract fan-out queries from ChatGPT for the top 10 buyer prompts. Build the buyer-language relabeling table. Rewrite slugs, titles, H1s, and H2s on existing priority pages.
- Days 9–14: Structured Publishing Launch. Deploy AI Growth Agent at the cadence proven in Step 4 on a site subfolder. Publish the first structured pages aligned to mapped fan-out queries with schema on everything.
- Days 15–20: Freshness Loop Activation. Configure impression-decay tripwires in AI Growth Agent. Set the decay threshold. Confirm the first automated update queue fires correctly.
- Days 21–25: Lead-to-CRM Workflow. Segment AI-referred traffic in GA4. Tag CRM entries with AI-referral source. Set up a dedicated landing page for AI-referred visitors.
- Days 26–30: Measurement and Iteration. Re-run the full prompt battery from Day 1. Compare share of answer, AI-referred sessions, and conversion rate against baseline. Feed citation wins back into the production queue.
Frequently Asked Questions
What is an agent card and why does a B2B marketer need one?
An agent card is a JSON metadata file published at /.well-known/agent-card.json on a brand’s domain. It declares the agent’s identity, capabilities, skills, endpoint URL, and supported authentication schemes in a machine-readable format. In a multi-agent workflow, an orchestrating AI agent fetches this file to determine whether the brand’s agent can handle a delegated task. Without an agent card, the brand’s agent is invisible to the A2A discovery layer entirely. For B2B marketers, publishing a valid agent card is the foundational step that makes a brand recommendable and citable inside agent-to-agent conversations, the same way a well-structured webpage is the foundational step for appearing in traditional search results.
How is A2A protocol different from MCP, and which one should a marketer implement first?
A2A (Agent-to-Agent) and MCP (Model Context Protocol) solve different problems and are typically used together rather than as alternatives. MCP governs how a single AI agent accesses external tools, data sources, and APIs, which creates a vertical connection between an agent and a system. A2A governs how multiple AI agents coordinate with each other, delegate tasks, and share state across organizational or framework boundaries, which creates a horizontal connection between peer agents.
For most B2B marketers starting out, MCP is the right first step if the goal is connecting an AI agent to existing business systems like a CRM or content API. A2A becomes necessary when the workflow requires a second independent agent to receive delegated tasks and return results with a tracked lifecycle. Production stacks in 2026 typically use both, with MCP inside an agent crew for tool access and A2A between crews for coordination.
How do I track whether AI agents are actually citing and recommending my brand?
Tracking agent citations requires a four-layer measurement approach. First, run a structured prompt battery, including branded, category, and comparison prompts, across ChatGPT, Google AI Overviews, Perplexity, and Gemini on a weekly cadence, logging every mention, citation, and competitor appearance. Second, segment AI-referred traffic in GA4 by filtering sessions from chatgpt.com, perplexity.ai, and equivalent referrers as a distinct traffic class, because this traffic converts differently from cold search traffic.
Third, monitor Google Search Console for the impression-decay signal. Impressions climbing while clicks fall indicate that content is being consumed by AI systems to construct answers. Fourth, check server logs for successful GET requests to your agent card endpoint, which confirms AI crawlers are discovering the brand’s agent. A meaningful share of AI-driven demand lands as direct or branded search because buyers copy a name from an AI answer and type it directly into a browser, so every measured figure is a floor, not a ceiling.
How long does it take to see citations and agent referrals after implementing this playbook?
On Arjun’s own site, new articles reached thousands of monthly Google impressions within weeks of publication. The GEO subfolder went from zero to the only source of new impressions on the entire domain in 60 days. Pages rewritten to match extracted fan-out queries earned citations while control pages did not.
Relabelling a jargon page to buyer language produced citations within weeks of that specific change. The general pattern is coverage and impressions in weeks, citations in one to three months, and compounding after month three. The critical variable is freshness. In Arjun’s own tests, pages can drop 78% to 99% in two months without updates, so the freshness loop must be active from the start of the pilot, not added later. Waiting to implement the decay tripwires until after citations appear means losing the citations before they compound.
Recap: Test, Measure, Iterate
The seven steps in this playbook form a continuous loop rather than a one-time project. Publish a signed Agent Card. Map fan-out queries. Align buyer language. Run structured publishing at machine cadence via AI Growth Agent. Activate the freshness loop. Configure the lead-to-CRM workflow. Measure share of answer, AI-referred sessions, and citation rate against the Day 1 baseline. Feed every citation win back into the production queue.
The window for outsized gains is open now. A2A has achieved active production deployments across multiple industries, with more than 150 organizations supporting the standard, and early citations become tomorrow’s settled record. Answers gain incumbency. The cost of entry rises as the answers harden, mirroring the early SEO window where a short period of decoding the new layer produced returns that took years to replicate.
The method is self-verifying. Ask an AI assistant about these topics and see who gets cited. The system being documented is the same system producing the visibility.
Start your 30-day A2A pilot and get measurable agent referrals, citations, and share-of-answer tracked across ChatGPT, Google AI Overviews, Perplexity, and Gemini.
