A Marketing Team's Work, as Software.
Fractl Agents takes the jobs an agency does — research, press, search, content, data and surveys, creative, monitoring, client reporting — and runs them end to end. An agent plans the job, does it with purpose-built tools and the client's own data, checks the result, delivers it with its sources, and asks before it spends.
"Output is the only thing that actually matters."
Counts from the live platform catalogs and production usage logs, 22 August 2026. Job costs are per-tool medians from the last 90 days of production — measured, not modeled.
Every Job Follows the Same Shape
Plan → do → check → deliver — and ask before it spends. Every job on the platform follows that shape.
Plan
Brief it like a colleague. Click "Plan it with me" and a draft plan stands before your first question is answered — every step, the tool it uses, what it produces.
Do
The agent runs the plan with purpose-built tools and the client's own data. Change it in one sentence — the plan rewrites itself in plain English.
Check
Numbers are checked against their sources. Infographics are fact-checked by a panel of models. A finding that can't be evidenced is set aside and reported as such.
Deliver
Results land in the chat and the client's folder — with sources attached, as working objects you can edit, not walls of text.
Ask First
Every paid step stops on a card for your click. That stop is enforced in code, not by asking the model nicely.
Four Parts Make an Agent — Configured, It Becomes a Play
The Specialist
e.g. the Ideation Strategist — trained on how Fractl analysts actually brainstorm: analogical thinking, recursive critique.
Focused Moves
Score · Novelty check · Analogical leap — moves the expert calls, reusable across other tasks.
The Procedure
Generate → evaluate → critique → regenerate. A recursive loop with sub-agent evaluators — first-pass ideas are never the best.
Optional Trigger
Fires when fresh industry news lands, or weekly — the same job on a schedule, delivered to the bell, email, or Slack.
Two Ways In — Same Primitives Underneath
Pick a Play
Click a Play and the chat opens already shaped to do it — the expert loaded, the workflow wired, the knowledge in scope. Output in minutes, the same shape every run, built on agency methods.
Open a Chat
Talk to the right expert directly and shape the work in plain English. Slower than a preset — but when it works, promote it to a Play and it runs the same way forever.
Three Whole Jobs, Start to Finish
"Turn this dataset into a campaign"
Defensible findings mined in SQL → the winning angle stress-tested → fact-checked infographic, headlines, sourced long-form draft → journalist dossiers and verified contacts → pitches sent from the analyst's own Gmail → weeks later, Campaign Footprint finds every placement. A media list in 42 seconds, pulled from a 138-article dataset the agent just built.
"Where do we stand, and what next?"
Add a client from a URL and the brand profile drafts itself. Site Intelligence crawls the whole site; Exhaustive Search Diagnosis runs ~140 checks on the client's own Search Console and Analytics; reviews, Reddit, backlinks, local rankings and AI visibility fill in the Cockpit. By the second month it already knows the client — so the same work costs less.
"Why isn't AI recommending us?"
AI Visibility Check measures how often each AI engine names the client versus rivals. AI Search X-Ray surfaces the hidden searches and cited pages. The Hallucination Audit finds what AI wrongly believes — then the fix: citation targets, unlinked mentions, comprehension rewrites. Thirty days later, the check runs again.
136 Purpose-Built Tools, Seven Marketing Families
Most "AI for marketing" is a thin layer on a model. These tools aren't. Each one has its own method, pulls grounded data, states what it's for and not for, and returns two tiers of output: a client-facing report and an analyst-facing one with the evidence.
SEO, GEO & Content Strategy
A whole-site crawler auditing ~20 areas with a 45-view report dashboard. Search Console + Analytics diagnosis running ~140 evidence-bound checks. The newest sub-family: how AI assistants see a brand — visibility, citations, comprehension, hallucinations.
PR, Media & Outreach
Journalist discovery with a pitch-ready dossier per reporter. Bulk contact enrichment with verified emails. Newsjack discovery with a standing monitor. Campaign Footprint traces every placement a campaign earned and how each outlet credited it.
Research, Web & Sources
Information Gain Finder surfaces what ranking pages miss versus what experts and forums actually discuss, with verbatim quotes. Ideation Stress Test scores ideas 0–100 across six failure modes. Idea Genome is grounded in ~1,000 high-performing pieces across 12 verticals.
Data, Statistics & Surveys
Public statistics from seven federal sources, graded for defensibility. Data Story Miner computes every number in SQL. Survey Builder writes a fielding-ready instrument, fields it, and analyzes responses for statistically defensible findings.
Creative Media, Audio & Video
Infographics fact-checked by a panel of models before they ship — with an accuracy confidence score that re-runs after every edit. One study becomes a full short-form package: carousel, reel, kinetic video, audiogram, two-host podcast.
Social Listening & Publishing
Search and standing monitors across X, Reddit and LinkedIn. B2B Reputation Scan across G2, Capterra, Trustpilot, TrustRadius and more. CMS publishing and social posting — each with preview and confirm.
Lead & Contact Enrichment
Find a person's work email, validate one you have, list a company's employees, find who holds a role, read a LinkedIn profile, enrich a company — any of those in bulk.
Connected to Where You Work
Search Console, Analytics, Drive, Slack, HubSpot, Salesforce, Google Ads, SurveyMonkey, Jira, Asana and more. The expensive SEO and journalist data runs on Fractl's account — so every user gets it.
Working Objects, Not Walls of Text
Results arrive as purpose-built cards: images with an editor, video with trim and captions, data tables that open into a workbench, reports that revise by instruction. Every chat keeps an outputs tray, downloadable as a zip.
What "a Real Instrument" Means
Exhaustive Search Diagnosis takes full Search Console and GA4 history through ~140 typed checks, and a judge must evidence every finding from the data — "a finding that can't be evidenced is set aside and reported as such." Four documents come out: executive brief, client report, evidence appendix, full internal report. Campaign Footprint searches the open web, two backlink indexes and news APIs, reads every candidate page and keeps the sentence that proves the match — pages it couldn't verify go to a human review queue rather than being dropped. Every other tool in the catalog is held to the same standard: something an analyst can stand behind.
Tools Are the Parts. The Layers Above Make the Work Repeatable.
Plays, experts, skills, workflows and automations compound into a catalog — and each layer has an AI wizard, so your team builds its own without touching anything technical.
A Play is a job with its steps, inputs, outputs and guidance, so the agent runs it the same way each time and a bar tracks the steps. Earn press coverage. Win organic search. Manage how AI sees you. Produce an original research study. Fact-check and QA content. Find and reach prospects — and thirteen more categories.
A team member with your method: its own instructions, tools, and the skills and workflows it may call. Build your own — name its niche and it runs an original research pass of about twenty-five minutes to learn what the base models don't know, compiled into a browsable knowledge wiki it can cite. Team-built experts are your IP, private to your workspace.
A move written once — instructions, tools, scripts, reference files — callable by any expert. Bring in skills built anywhere in the Claude ecosystem: one import handles a single skill file, a zipped package, or a whole archive.
Some work should not be improvised. Judgment gets an agent step; certainty gets a fixed step; a person gets a human checkpoint — in five forms, from approve/reject to pick-rows-from-a-table. Describe the process in plain English and the builder lays out the canvas; a builder agent checks run-readiness and fixes the graph on request, with undo after every change. If step seven fails, steps one through six are kept.
A saved request, a workflow, an X monitor, a newsjack monitor, a competitor watch, a LinkedIn brief — daily, weekly or monthly, delivered to the bell, email or Slack. Each carries rules for how the chat agent may use its results.
Everything It Does Is for a Specific Client — and It Gets Better With Use
Add a client from a URL and the brand profile drafts itself from the website. From then on, every conversation, file, deliverable, meeting and decision is filed to that client — walled off from every other client.
Brand Profile & Voice
Around thirty fields drafted by AI, corrected by you: audience, competitors, approved claims and stats, things not to say, goals, target publications, brand kit. Writing styles are built from your own sample articles and assigned per use case — email, LinkedIn, press pitch, blog, executive memo.
Knowledge & Memory
Built from the client's email threads, Slack mentions and Drive documents — each fact carrying a trust badge. A living dossier holds rules, decisions, and what didn't work; a review lane confirms or corrects facts; the dossier exports as a client brief.
The Fractl Notetaker
Connect a calendar and turn recording on per client. It joins Google Meet, Zoom or Teams, reads a consent line when it joins, and files notes, recap and action items to the client. Ask afterward what was agreed.
The Cockpit
A per-client dashboard across eight areas — organic, content, AI visibility, PR, links, technical, local, reviews — with ranked findings and a button on each to run, re-run, or take into chat. Thirty-two instruments feed it, and a settings panel shows what each page is built on, last run and cost, and what it inferred versus what you declared.
Every Deliverable Carries Its Receipts
Lineage
Which expert ran it, which skills, which workflow steps, how many tool calls — re-runnable on new inputs.
Sources
Every claim tied to a dataset, news article, internal file, or prior deliverable. Click a chip to see the cited passage.
Confidence
Auto-verification scores sourcing, novelty against the archive, and rubric fit. Anything under 0.7 is held back for human review before saving.Verified · 0.91
Most Tools Stop at a Report in a Folder. This One Finishes the Job.
It sends the pitch, publishes the post, puts the page on a client site, and watches the market on a schedule — with a confirm step before anything leaves.
Outreach From Your Own Gmail
A campaign workspace ties a list to the specific thing pitched, for a specific client. Pitches drafted per person — beat, recent work, the client's voice — sent from the analyst's own Gmail on a send window and cadence, so replies come back to a human being, not a platform address. Replies triaged, follow-ups scheduled, coverage logged as it lands. You see the list and the copy, and nothing leaves without your yes.
Publish to the CMS
Drafts into WordPress, Ghost, Drupal, Strapi, Directus and Joomla — connect a destination, map fields, test. Create or update drafts, attach images, preview, schedule; pull existing posts into a table for a bulk refresh. Each publish has a confirm step; live and scheduled publishing require separate approval.
Post & Share, Sign-Off Built In
LinkedIn, Facebook, Reddit, Threads, Bluesky, Google Business — preview and confirm before anything goes out. Clients sign off via a zero-login review link: comment paragraph by paragraph, approve. Going back to a last published version is one click.
One Roster, Standing Watch
Journalists from media search, pitch analysis and imported lists land in one shared team roster — beat, outlet, contact, which campaigns they've been in. Contact enrichment pays wave by wave, only until it has the verified emails you asked for. Monitors for news hooks, X, LinkedIn, competitors and site changes report to email or Slack.
Teams Abandon Marketing AI When They Stop Trusting It
Long before they run out of uses for it. So trust lives in three checkable places — not in a promise.
In the Outputs
Numbers in research drafts are checked against their source and flagged when they can't be. Infographics are fact-checked by a panel of models. Every tool states what it's not for — and reports when it couldn't verify, rather than guessing.
In the Approvals
Paid data, emails, posts and publishing stop at a card you click — enforced in code, not by asking the model nicely. More than twenty kinds of card, from approve and rank to type-to-confirm the irreversible. Stop cancels the run and any waiting card.
In the Accounting
Every run metered per user and per client; every action logged with an admin audit log. Failures are captured automatically and triaged nightly. Cost overview with CSV export; an evaluation harness drives the real chat in a test environment before changes ship.
Cost of Goods Is Measured, Not Modeled
Every tool call and model token is logged with its cost, by user and by client. Total compute for the whole agency over the last 30 days: $1.2k — $819 in model tokens, $350 in paid tool calls. Average compute per conversation: $1.90. Most single tool calls cost cents.
| Job | Typical | Heavy run | Estimated replacement |
|---|---|---|---|
| AI hallucination audit — 6 models, accuracy grade, false claims, the fix | $0.04 | $0.04 | A day of querying with no grade or record |
| Media list — journalists on a beat, dated articles, fit rationale | $0.01 | $0.05 | 2–3 analyst hours |
| Search Console / GA4 diagnosis — ~140 evidence-bound checks | $0.55 | $7.20 | A week of an SEO analyst |
| Original research sprint — public stats, mined communities, expert quotes | $1.40 | $1.60 | Days of analyst time |
| Long-form draft — every number checked against its source | $6.20 | $5.60 | A writer-day plus an editor pass |
| Data story miner — rankings, trends, correlations in SQL | $0.50 | $0.50 | An analyst-day |
| Infographic — rendered, then fact-checked by a model panel | $0.30 | $0.30 | Designer plus analyst, a day |
| AI Search X-Ray — hidden searches, 200+ pages scored cited vs. skipped | $0.50 | $0.55 | Not possible by hand |
Eight of the eleven jobs itemized in the partner briefing (which itself shows 11 of 50). "Heavy" is the average when a run goes deep; the replacement column is Fractl's estimate. Pricing is being set from real usage.
Fifteen Years of Agency Method, Written Down and Run
Fractl has done SEO, content and digital PR for fifteen years. The people who wrote these Plays are the people who pitch journalists, audit sites and run studies for clients — and they judge the outputs the way they always have: would this go to a client?
A software company building "AI for marketing" has to guess what a newsworthy statistic looks like. Here the answer was already in the building — and the platform runs on real client work, so the rough edges show up the same week and get fixed. The AI makes the work faster; it didn't change what counts as client-ready.
5,000+ Campaigns Delivered
Founded 2012 in Delray Beach, FL. Campaigns for Adobe, Discover, Upwork, Paychex, AutoNation, DIRECTV, ADT, Indeed, McAfee, Care.com, Porch and 200 more. The Plays were built and tested on this work, not on synthetic data.
Research the Press Cites
Fractl's studies have been featured in The New York Times, Harvard Business Review, TIME, Fast Company, Forbes, Inc., USA Today, The Washington Post, NPR, CBS, NBC, MIT Tech Review, Entrepreneur and the trade press.
Results Clients Report
The Porch partnership added over 23,000 new monthly site visitors in one year. One case study saw a 1,100% organic search traffic increase over 36 months. HeliosX earned 500+ links across both brands with 251% year-on-year SEO traffic growth.
Kelsey Libert
Co-founder, SVP Marketing — four Harvard Business Review bylines; MozCon, BrightonSEO and SMX speaker; Search Engine Land contributor.
Kristin Tynski
Co-founder, SVP Research — quoted in The New York Times on viral content; BrightonSEO speaker on AI and earned media.
Nick Santillo
CEO — 20+ years; architect of the agency's Plays methodology.
Dan Tynski
SVP Technology — built the Fractl Agents platform; 15+ years engineering content and data products.
Straight Answers to the Hard Questions
Isn't this just ChatGPT with a nicer UI?
Am I locked in?
What about hallucination?
Where does my data go?
Can I make it work my way?
Does Fractl actually use this?
Do I need someone technical to set it up?
Don't Go Looking for Tools. Bring One Company You Care About.
Tell the agent the goal, plan the work with it, and watch what it does.
Log in and let onboarding set you up
Your team, then your first client — name and website, and the brand profile drafts itself in about thirty seconds. A few minutes in, the system is already about your client, not a blank chatbot.
Plan it with the agent
Click "Plan it with me" and brief it like a strategist: "We need press for this launch." "Why isn't the site getting the traffic it should?" "There's a story in our own numbers somewhere." You never name a tool — and the first paid step stops on a card.
Push it forward
Say yes to spend. Results land finding-first; click the next offered step or say what you want. Type a note while it works and it folds it in. Each step hands real data to the next.
See what it remembered
A second conversation starts where the first ended — decisions, preferences, what didn't work. Walled off from every other client.
Build something of your own
Open Build in the header. Every builder has a wizard; a new expert researches its own knowledge wiki while you watch. Nothing in there is magic.
Fractl Agents is live at chat.frac.tl, in daily use at Fractl and with the first client teams. Accounts start with a balance — when it asks permission to spend, say yes.