Why Quarry Runs Outbound And AI Visibility As One System

Quarry is an agentic AI demand engine I built to run outbound prospecting and AI visibility as one loop instead of two separate jobs. A crew of specialist AI agents researches accounts, drafts outreach, and works on how a company appears when a buyer’s own AI researches it back, with a human approving every send.

I started Quarry after watching the same pattern play out across list-based outbound tools: high volume, low reply rates, and no connection to what happened when a prospect later typed the company’s name into ChatGPT or Perplexity. Outbound and AI visibility used to sit in separate budgets, run by separate teams. I built Quarry on the argument that split stopped making sense once buyers began forming opinions before a seller ever got a reply.

Why did outbound stop converting?

Outbound stopped converting because the reply pool it depends on has been shrinking for years, not because sales teams got worse at writing emails. Average cold email reply rates fell from around 8.5 percent in 2019 to about 3.4 percent by 2026, according to Amplemarket’s 2026 cold email benchmark data. Buyers now do most of their research before a seller hears from them at all: Gartner’s own research finds B2B buyers spend only 17 percent of their purchase journey meeting with potential suppliers, split across a buying group that Gartner puts at six to ten stakeholders who frequently disagree with each other. A generic message sent to a scraped list, arriving mid-way through that process, competes for attention it was never built to win.

Why merge search and outreach?

I merge search visibility and outreach because both now serve the same moment: the point where a buyer decides who to trust, whether that happens on a call or inside a chatbot. Even before AI assistants entered the picture, most Google searches already ended without a click, according to SparkToro’s 2026 analysis, which found under a third of searches still send traffic to a website. Search Engine Land’s 2025 tracking found that when an AI Overview appears on a query, click-through rates on the underlying links drop by close to 60 percent. If part of a buyer’s vetting now happens inside an AI assistant rather than a phone call, how a company reads to that assistant is part of the pitch, not a separate marketing project running on its own budget.

What is a specialist agent crew?

A specialist agent crew is a set of separate AI agents, each built for one narrow step, coordinated through a shared process instead of one general model attempting every step itself. Quarry works this way: one agent researches an account, another drafts a message, another checks it against deliverability rules, and a coordinating layer keeps their output consistent. Anthropic’s engineering team described a similar tradeoff in its own research tool, where multiple agents working in parallel covered more ground than a single agent chain, at the cost of using several times more tokens. IBM draws the underlying distinction as agentic AI pursuing a goal with limited supervision, versus generative AI that only responds to one prompt at a time. Running several coordinated agents costs more and is harder to debug than a single script, which is one reason most outbound tools have not built one this way.

The five checkpoints, ranked by cost of skipping

These are the five checkpoints Quarry’s agent crew runs a prospect through, ranked by how quickly skipping each one kills a reply:

1. A verified trigger event, not a static list. Sending because something specific changed for that account beats sending because a scraped list had a matching job title.
2. A message written for that account, not a template with fields swapped in. The research agent’s findings have to change the draft itself, not just the greeting line.
3. An authenticated, warmed sending domain. Google’s bulk sender rules, in force since February 2024 for any domain sending 5,000 or more daily messages, require SPF, DKIM and DMARC before a message is even scored for relevance.
4. A visibility layer consistent with the outreach message. If a buyer checks the company out mid-conversation and finds something that contradicts the email, the email’s credibility drops with it.
5. A human reviewing the send. Every draft above this point is a proposal until a person approves it; nothing goes out on its own.

Why does a human have to approve every send?

A human has to approve every send because the downside of a bad automated message is asymmetric: one good email ignored costs nothing, one bad email sent at scale can get a domain blocklisted for months. The FTC’s CAN-SPAM rules make the sender legally responsible for a commercial email’s content and opt-out mechanism, regardless of whether a person or a tool drafted it. Deliverability data backs up why that caution is warranted: Validity’s 2025 Email Deliverability Benchmark Report put global inbox placement at around 84 percent, with new, unwarmed domains reaching the inbox as little as 12 percent of the time. Gartner has separately warned that as agentic systems take on more autonomous action, organizations need explicit control over how much autonomy those systems get. Gating every send through a person is slower than full automation. That tradeoff is deliberate: this is architecture, not a setting a customer could switch off.

Old outbound and SEO vs. a coordinated loop

Old-style outbound and SEO differ from a coordinated agent loop on four criteria that decide whether a message reaches someone already inclined to reply: verification, availability to AI systems, trust, and machine readability. None of this makes the old approach obsolete. It is cheaper to start, needs no agent coordination, and can still move a lot of volume fast when a list is genuinely warm. What changes is which parts of the job get checked before a message goes out, and whether a company’s own site is written for a person skimming it or a model parsing it.

• Verification. Old way: a list is judged by open rate after the fact. New way: an account is confirmed against a specific, checkable event before the first draft is written.
• Availability to AI. Old way: content is written for search rankings and rarely checked against what a chatbot says about the company. New way: the same facts a salesperson would say out loud are kept consistent with what an AI says when asked; HubSpot has reported its own organic traffic fell 27 percent year over year even as AI-referred visits convert at a higher rate.
• Trust. Old way: reputation builds from repeated exposure across months of campaigns. New way: one accurate, well-timed message can substitute for repetition, but only if the sending domain is authenticated and the claims are accurate when checked.
• Machine readability. Old way: a brochure site written in marketing prose a language model has to interpret rather than read directly. New way: plainly stated, consistent facts a model can quote without guessing at what a paragraph of adjectives meant.

Where does this approach struggle?

This approach struggles most where the underlying facts are thin or where the market cannot yet be measured by any outside system. If a company has no distinct case for why a buyer should choose it, no amount of well-timed research or accurate sends changes that outcome; Quarry can find the right moment to speak, not manufacture a reason to be believed. AI visibility is also a genuinely unstable target to optimize for. Gartner’s own 2024 forecast that search volume would fall 25 percent by 2026 has not played out as projected: Search Engine Land reported in January 2026 that U.S. organic search traffic, based on Graphite’s analysis, was down only about 2.5 percent year over year. A visibility approach tuned to how today’s AI models cite companies may need retuning as those models change, and there is no way to know in advance by how much.

Who is this not for?

Quarry is not built for a company that wants to send the largest possible volume of near-identical messages with no one checking them; the human-approval step is a deliberate ceiling on speed, not something to route around. It is also not a fit for an idea-stage product without paying customers, consistent case studies, or a clear point of view yet, because there is nothing distinct for the visibility side of the loop to reinforce. And it is not a replacement for a sales team’s own judgment on qualification or pricing: Quarry decides who to approach and how the company should read when checked, not what to sell them or for how much.

You are a B2B go-to-market analyst. I'm going to describe my company and I want you to tell me what a buyer's AI assistant would likely say about us if asked to compare us to competitors, based on general patterns of what LLMs tend to surface (recency of content, third-party mentions, structured facts versus marketing prose, consistency across sources). Here is my company: [describe product, ICP, main competitors, and 3 to 5 factual claims about what you do]. Give me: (1) what's likely easy for an AI to find and cite accurately, (2) what's probably missing or inconsistent, (3) three concrete changes to make our facts easier for a model to quote correctly. Ask me clarifying questions first if you need more detail about my market or competitors.

You are a sales research assistant. I'll give you a company name and its public website. Research and summarize: what specific, recent trigger event (funding, leadership change, product launch, hiring surge, public complaint) would make this account plausibly ready to talk to us right now, versus generic reasons that apply to any company in the industry. Company: [name and URL]. What we sell: [one paragraph]. Return a short brief: the single best trigger you found with a source, two backup angles, and one open question you couldn't verify. If you can't find a genuine recent trigger, say so directly instead of inventing one.

You are a deliverability and compliance reviewer for outbound email. I'm going to paste a draft cold email below. Check it against: CAN-SPAM requirements (accurate from and reply-to fields, no deceptive subject line, working opt-out), whether the claims made about my company are ones I can back up, and whether the tone reads as personalized to this specific account rather than templated. Draft: [paste email]. Context on the recipient and why I'm reaching out: [paste]. Return a pass or fail on each check, and rewrite only the parts that fail.