How we built an AI agent for website optimization, SEO and Google Ads

Every project we run gets an agent that reads their website the way a senior marketer would: where visitors come from, what they do, what it costs, what broke. Then it proposes the one to three changes that matter, and a person makes them. Here is the setup, on our own product first.

Jaro16 min read
Summary
Share
The 3DAY Agent mark: dark horizontal lines forming a rounded shape, on the brand red

Key takeaways

  1. 01It evaluates a site in the order a buyer moves: traffic, behavior, cost, health.
  2. 02Code computes every number. The model only ranks, explains and drafts.
  3. 03Every report opens with the same overview, so each period compares with the last.
  4. 04It lists what shipped on the site, so a drop comes with its likely cause.
  5. 05Nothing changes from Slack: a person makes every change, by the improvement's ID.
  6. 06Between reports it checks the site hourly and the analytics 3 times a day.

Why we built it

Running a website well means reading Google Ads, Search Console, analytics, the code and the inbox together, every week. We do that for the projects we run. The signals that matter most sit in the gaps between those tools: a paid search that lands on a page that does not convert, a signup step that broke after a deploy, pages that slowly slip in Google. Finding them takes a senior person about an hour per project, before any decision is made.

So we built an agent that does that reading every week or two, the way our team does it, and comes back with the one to three things that actually matter. Not a dashboard, not a score: a short note in Slack, with fixes ready to approve. Our people spend their hour on the decisions instead of the exports.

We are building it on our own product first, InstantSchema, before it touches a client site.

What it is connected to

The agent reads nine sources, thinks with Claude and one memory, and talks in three places. The map shows what runs today.

Google Adscampaigns, search terms, costSearch Consoleclicks, impressions, positionsDataForSEOGoogle results, links, reviewsPostHogfunnels, errors, recordingsStripepayments, customers, MRRLive siteloads every hourSite codewhat shipped, GitHubEmaildelivery or list, repliesMentionsnews, Hacker News, web search3DAY AgentClaudedrafts words, never numbersBrainshared rules, in codeProject specone project's rulesNeon Postgresdata, items, alerts#project in Slackthe report, on its cadence#agent in Slackalerts and reviews, privateapp.3day.studioreports, items, settingsAdsSearchSERPPostHogStripeSiteCodeEmailNewsAgentClaudeBrainSpecMemorySlack#agentApp

The agent

3DAY Agent

  • Pulls every source and computes every number in code
  • Checks key numbers through a second path
  • Checks the data can be trusted before any item: broken tracking, silent funnel steps, payments that stopped, late data
  • Ranks findings by money at stake and writes up to 3 items
  • For a site under 100 visitors a week: a launch checklist built from its data, and up to 3 items from it
  • Under the items, one piece of content worth making with its kind; every other idea in the app
  • Changes nothing on its own

Every hour for health, Monday 08:00 local for the report

  • Google Ads (Reads, read-only): Spend, clicks and results per period; The search terms people actually typed, with cost and results; Every change in the account, with old and new values and who made it. Every report, Monday 08:00 in the project's timezone.
  • Search Console (Reads, read-only): Visitors from Google, times shown and average place; The searches people use and the ones climbing, with visits from searches for the brand as one row; Search chances: close to the top, rarely clicked, article ideas; For a young site: which main pages Google has indexed, page by page. Every report, with Google's two to three day lag.
  • DataForSEO (Reads, read-only): What Google shows for what the project sells: places, AI answers, forums, videos; Searches it could win, with volume and how hard the first page is; Lists, threads and directories to get onto, and which ones Google's AI answer cites; Reviews, and the sites that link to competitors but not to the project; Only for projects switched on; one pull a week, within 1 USD. Every report, from one weekly pull.
  • PostHog (Reads, read-only): Every step of the funnel, from first visit to the main goal; Rage clicks and errors; One recording worth watching per report; Visitors from AI assistants (ChatGPT, Perplexity, Claude, Gemini) and whether they convert; Health check: page views, errors and the main goal against last week. Every report, and health checks at 06:00, 12:00 and 17:00.
  • Stripe (Reads, read-only): Money received, orders and new paying customers per period; MRR by Stripe's own definition, against last period and a month ago as reported then, and subscriptions past due; What went wrong, cause by cause: failed payments and where they stand, refunds, disputes, cancellations, abandoned checkouts; A restricted key with Read only; the code makes read requests and nothing else. Every report.
  • Live site (Reads, read-only): Loads the home page every hour; Down only when it fails twice, 20 seconds apart; Bot protection (401, 403, 429) counts as up; For a young site: each main page's title, description and share image. Every hour.
  • Site code (Reads, read-only): Every change that went live, with its date; Flags changes to tracking, sign-up, payment and SEO; Marks them on the PostHog charts. Every report.
  • Email (Reads, read-only): The project's email tool, Resend or Flodesk, in the same block; Resend: what the emails did, delivered, bounced, marked as spam, clicked, unsubscribed; Flodesk: subscribers at the end of the period and new ones; what Flodesk does not give shows n/a; Replies and new emails from people, read from the project's inbox, sender hidden; Up to three worth reading: might leave, something broke, wants to buy, praise, a request. Every report.
  • Mentions (Reads, read-only): Google News and Hacker News; One web search per platform: X, LinkedIn, YouTube, Product Hunt, G2 and more; Only links that came back from a source, each shown once; For a young site: a checklist fixed by business type, with the review platforms for its type. Every report.
  • 3DAY Agent (The agent): Pulls every source and computes every number in code; Checks key numbers through a second path; Checks the data can be trusted before any item: broken tracking, silent funnel steps, payments that stopped, late data; Ranks findings by money at stake and writes up to 3 items; For a site under 100 visitors a week: a launch checklist built from its data, and up to 3 items from it; Under the items, one piece of content worth making with its kind; every other idea in the app; Changes nothing on its own. Every hour for health, Monday 08:00 local for the report.
  • Claude (The agent): Ranks the findings and explains them in plain words; Picks the real mentions from the search results; Sorts the emails people wrote into what needs a person; Sorts replies under the reports: a project rule, a Brain change, a question or a request; Never writes a number: a draft with one the code did not compute is rejected. A few calls per report.
  • Brain (The agent): The checks every project runs, with their thresholds; Data checks first: no item from tracking that broke or data that came late; Never proposes blocking a project's own keywords or its brand; Changes go through code, tests and How it works, for every project at once. Every run.
  • Project spec (The agent): The project's sources, goal, funnel and report frequency; Its own exceptions, shown on the Setup tab; Every change is a new version, with who and why; Rules from client replies: searches never to block, checks switched off; Notes from replies the model reads as context. Read at every run.
  • Neon Postgres (The agent): Every pull, number and report; Every item and what happened to it: open, applied, not doing it or expired; Open health alerts and the mentions already shown. Always, the agent's memory.
  • #project in Slack (Talks to): The report, every week or every 2 weeks; Items with IDs: reply with an ID and the team acts on it; Replies read every hour, sorted and proposed to #agent; a person applies them in the admin. Monday 08:00 in the project's timezone.
  • #agent in Slack (Talks to): Health alerts: site down, no page views, errors, no main goal; New projects' reports, until they are live; Missing data, AI cost alerts and the weekly cost report; A person decides what the client is told. The hour a problem is found.
  • app.3day.studio (Talks to): Every report with its numbers; Items to mark done or not doing, and the log; Every content idea, ranked, with its kind and searches a month; Settings: report frequency, timezone and customer value; Replies to apply or dismiss, each project's spec, and the Brain changes waiting. Always on.

The agent in 14 slides

What a report shows, section by section, as the dashboard draws it. To keep or to send to your team.

Download PDF, 6 MB

How it evaluates a website

The agent follows the same order we use when we evaluate a site by hand, which is the order a buyer moves through it: how they arrive, what they do, what that costs, and whether anything is in their way. A report reads in five chapters: the overview and the money, organic traffic (AI assistants and Google), paid ads, conversion, then mentions and emails. What to do comes right after the first chapter, so the first screen holds the numbers that matter and the actions.

1. Traffic: where visitors come from

First the two doors into the site. Paid traffic comes from the Google Ads API: campaigns, the search terms people actually typed, cost and what each click led to. Organic traffic comes from the Search Console API: clicks, impressions and positions per search and per page. For projects where we switch it on, DataForSEO adds what Google itself shows for what the project sells, once a week: the searches it could win, the lists, threads and directories that rank and which ones Google’s AI answer cites, its reviews, and the sites that link to its competitors but not to it.

Reading both side by side matters because they behave differently. Paid traffic stops the day the budget does and every click has a price; organic traffic builds slowly and is free. The agent watches how the mix shifts week to week, because that decides where a wrong number costs the most.

Paid traffic: a Search campaign over 30 days, with clicks, impressions, average cost per click and spend.
Organic traffic: four weeks of clicks and impressions from Google Search Console.

2. Search terms: what people actually want

Keywords are what we bid on; search terms are what people actually typed before clicking an ad. The agent reads the search terms, because that is where money leaks: it flags the ones that cost money and led to nothing, and shows the keywords that bring the most conversions.

Every report shows the three searches that brought the most visits, with their place on Google, how often they were shown and the visits, and up to three climbing ones: new searches, or ones that climbed five places or more. On the paid side it shows the three keywords with the most conversions and up to three climbing ones, with clicks, conversions and the cost of each conversion. Like every number in the report, it compares the report’s own period with the one before: a week against a week, or two weeks against two.

Keywords with clicks, conversions and cost per conversion. The agent shows the three with the most conversions and the ones that are climbing.

3. Behavior: what visitors do

Traffic means nothing until you see what happens next, so the agent reads the product funnel in PostHog: every step of the project’s funnel, from the first page view to the main goal and past it where the funnel goes further, how many people make each step, how many leave at each one, and what share of the first visits gets that far.

A six-step product funnel in PostHog, from the first page view to a converted trial, with the drop-off and time between each step.

The agent looks for the step with the biggest drop that the site can actually influence, and that is where its proposals point.

4. Cost per real outcome

This is where joining sources pays off. Google Ads reports conversions, which often look healthy. But a conversion is only what the account was told to count, and that is often an early step like a started form, not a signup or a payment. So where a project’s Ads conversion is an early step, the report names it for what it is, like “generations started”. It divides Ads spend by the main goal PostHog recorded from Google Ads visitors, and with a profit setup, by paying customers too. The real cost per outcome is usually several times the number Google Ads shows.

5. Revenue: what came in and what went wrong

Where a project takes payments through Stripe, the second block of every report is money: revenue received with the number of orders, monthly recurring revenue against the same day a month before (the value a report showed then, since Stripe keeps no history and a past month worked out today can differ), new paying customers, new and canceled subscriptions, subscriptions whose payment is overdue, trials that converted, failed payments and refunds. Each count sits next to its money, because three cancellations of the cheapest plan and three of the biggest are not the same news. Recurring revenue follows Stripe’s own definition, so it reads the same as the client’s Stripe dashboard. A shop that sells one-off items sees only the rows that apply to it.

Under the numbers, a second table lists what went wrong, one cause per row, most money first: a card declined for lack of funds that Stripe is still retrying, a payment that needed 3D Secure and was never confirmed, a refund asked for within a day of paying, a dispute with the date it must be answered by, a cancellation in the first month, a checkout left unpaid. Each row says in one line what it means, from a fixed list in code, and links straight to it in Stripe. When payments stop while visitors keep coming, the report says checkout may be broken instead of reporting a quiet week.

6. Health and context

Last, what could be in the way. Some problems should not wait for Monday, so the agent also watches between reports: it loads the site every hour and watches its certificate, and at 06:00, 12:00 and 17:00 in the client’s timezone it checks the analytics for missing page views, an error spike, a new error in the site’s own code or a day without the main goal. Each check compares the same hours a week earlier, so a quiet night never sets it off, and errors that come from in-app browsers rather than the site are left out. Anything it finds goes to a person, once, with what is behind it underneath: the cause it found, the page, the first thing to fix and a link to the evidence. It reminds a day later if the problem lasts and notes when it is back to normal.

Then the outside world: what people said about the project since the last report, from Google News, Hacker News and one search per platform, such as X, LinkedIn, YouTube, Product Hunt and review sites like G2. The model picks the posts that are really about the project and writes one line on each, but it never supplies a link: code keeps a mention only when its link came back from a source. Context never becomes a number in the report. It only explains one.

7. Email: what was sent and what came back

Most projects send email: sign-up confirmations, trial reminders, receipts, newsletters. If those stop arriving, nothing else in the report shows it. So the agent reads what the project’s emails did from Resend, the service that sends them: how many were delivered, bounced, marked as spam, clicked and unsubscribed, against the period before. Opens are shown too, labeled rough, because Apple Mail opens every email it receives. A project that sends newsletters through Flodesk gets the same block from its list instead: subscribers at the end of the period and how many joined. Flodesk does not report deliveries, opens or clicks, so those rows say n/a rather than showing a made-up zero.

Then it reads what people wrote back. The project’s own inbox holds replies to those emails and new emails from customers, leads and partners, mixed in with notifications and newsletters. Code keeps only emails written by a person, from outside the project’s domain. The sender’s address and name are never sent to the model, and addresses, phone numbers and links in the text are blanked first. The model sorts what is left: someone might leave, something broke, someone wants to buy, praise worth quoting, or a feature request. The report shows up to three, each with one line on what they want and a short quote, and code keeps a quote only when it appears word for word in the email.

8. What changed

Numbers rarely move on their own. The agent lists every change that went live in the period, straight from the site’s code, and every change made in the Google Ads account, from Google’s own change history, with the old and new value (a budget going from 30 to 50 a day), the keywords added or removed, and who made it, including recommendations Google applied by itself. The team’s review copy also shows every site change as written. Code flags the changes that touch what the numbers measure: tracking, signup, checkout and SEO. Those are also added to the PostHog charts, so a drop shows the deploy next to it. When a finding lines up with one, the report says so with the date: “since the new signup flow shipped on 9 October”. It says “since”, never “because”; the person reading decides.

If a site is not in code the agent can read, for example on Webflow or Wix, it cannot see what changed, and every few months the report can suggest what a custom stack would add. A project that runs no ads gets one line saying what Google Ads would show and how to ask for it, unless it told us it doesn’t use them.

What lands in Slack

Every Monday at 08:00 in the client’s timezone, weekly or every two weeks, the report posts to the project’s Slack channel. Every section shows its numbers the same way, so each period compares directly with the last:

  • the project’s own metrics, each with its change against the period before in its own column, in percent, as a count when the numbers are small (12 to 17 is +5), or in places and points for position and rates;
  • a direction for each one, because for average position or cost per signup lower is better;
  • the absolute numbers always sit next to the percent, so 1 to 5 clicks does not read as a 400% win.

It opens with the same overview, in the same order every week, each number with its change against the previous week and against the previous 4-week period. The main goal and its conversion rate, the same as in the funnel, the money received and the visitors, and when the project runs Google Ads, what a click costs, what a paying customer costs (CAC) and what a customer is worth against that cost (LTV:CAC). Everything else, from Google search to emails, is in the chapters below it. Right after that come at most three items, each with an ID by category, like CRO-2 for conversion, SEO-3 for search or REACH-1 for work beyond the site, what happened, why it matters in money or customers, and one proposed action. Under them, apart from them, one piece of content worth making, with its kind: a page for a search with buying intent that nothing on the site answers, an answer to a question buyers ask in a thread that ranks on Google, a study with original numbers, or, when the site already has a page for every search worth working on, the page to make the best answer for one of them. The app keeps every other idea, ranked. Then the details, in five chapters: the conversion funnel, a visit worth watching and what changed on the site; revenue; where free visitors come from (Google, searches for the brand name, AI assistants), the searches the site is found for with their place and visits, and the pages worth adding or improving for free visitors (from the site’s own searches: which page to improve, retitle or create, with the clicks it could bring); Google Ads with its cost per click, CAC and LTV:CAC and its best keywords; and last, mentions and the emails that went out and the ones worth reading. The numbers are real tables, so they line up on a phone too. A source the project does not track shows one line saying what it would add, and how to ask for it. Slack gets the short version: a long table, like what went wrong with payments or what changed in Google Ads, is one line there with a link, and the app shows the same line with the full table under it. Most days nothing else is posted.

When a site is too young to measure

A new site with 30 visitors a week has numbers that don’t hold still. A change has to be close to double before it is more than chance, and a conversion rate can’t be told from one three times as high. Reporting trends on that would be noise, so a site under 100 visitors a week gets a launch phase instead.

Every week or two it gets a few counts with their change against the period before (none under 20, where a swing is noise), a checklist right under them, up to three improvements the owner can make and one piece of content worth making, what Google shows about it, and its Google Ads once it runs them (or a first test with the searches and click prices to start from). The checklist is the foundation, fixed by the kind of business, so it never shifts with what one week’s data happened to show: every main page in Google, a title, description and share image on every page, structured data, the main goal tracked, which the agent ticks itself and the owner sees as one row, our post-launch check; then first reviews on the platforms buyers of that kind read (Google and Clutch for a studio, G2 and Capterra for software), reviews shown on the site, the directories that rank for its main search and the forums where buyers ask. The improvements come from that checklist and from what would bring visitors, each with a named target and a range: first a page worth adding for a search Google already shows the site for, with the clicks it could bring; then the owner’s part of the checklist, like collecting first reviews; only then ideas from Google’s results, such as a list of the best that Google’s AI answer cites or a recent thread that ranks for what the business sells. The phase ends when the checklist is done or visits stay at 100 a week for three weeks. The project then moves to its scaling phase and gets the full report, and the checklist stays as the record of what was done.

A new client’s reports go to our private channel first. We read them, fix what reads wrong, and only then switch the project live and let the reports reach the client.

The agent's first message in a project channel, the morning it was connected.

Approval, safety and memory

Nothing changes from Slack. Slack is where the agent reports and proposes; nothing anyone writes there changes anything. To act on a proposal, one of us opens a working session with the agent’s tools, names the ID, shapes the exact change, and applies it. The previous state is noted so the change can be undone, and the item is marked done in the dashboard.

Crucial problems go to a person first. A site down, missing page views, an error spike, a new error or a day without sign-ups go to a private channel with the cause the agent found, and a person decides what the client is told. A daily check for runaway Ads spend is next.

It checks the data before trusting it. Before any check runs, the agent asks whether the data behind it holds up this week. When Google Ads stops counting conversions while the clicks go on, or a funnel step suddenly records nothing, the report says the tracking broke instead of drawing conclusions from it, and checks that read late or missing data wait for the next run.

The model never writes the numbers. Code pulls every value and computes every change. A second check re-reads the key numbers, Ads spend, clicks from Google, revenue, and emails delivered or a list’s subscriber totals, another way, and the report says how many matched. The language model ranks the findings, explains them and drafts the text, and code rejects any draft with a number it did not compute, a link, a mention, or more than three items.

It remembers. Every pull, number, report, item, alert and mention is stored in a Postgres database in Frankfurt, so every number traces back to what the source said, and no mention is shown twice. It also reads every reply under its reports. Each hour it sorts new replies: a fact about this project (“those searches are our customers”), a check that is wrong for everyone, a question, or a request to act on an item. It posts each one to our private channel with what it would change. A person reads it and applies it or not. A fact goes into that project’s own rules, so the next report stops proposing what the client ruled out, and a flaw in a check becomes a fix in the agent’s code for every project. Once it is applied, the agent answers in the thread. Replies never change a site or an account, and never loosen the safety rules.

Access is read-only wherever possible. The scheduled jobs read Google through an account that can only read, read code through a read-only GitHub App, read the project’s inbox through a read-only Google permission the client’s admin grants and can revoke (Google cannot limit it to one mailbox, so the code reads only that one and never sends, labels or deletes), read Stripe with a restricted key that can only read and never a secret key that could charge or refund, read email stats with a key the code allows only to read, and log in to the database with a role that cannot change projects, limits or cost records. The account that can change Ads is used only by a person, when a change is made. Our keys live in encrypted secrets, never in code. A key a client pastes into the dashboard is locked with our public key before it reaches the database, and only the agent can open it. A check before every commit refuses anything that looks like a key.

Where it runs

One job runs every hour on GitHub Actions, started at a fixed minute by a small Cloudflare Worker, because GitHub’s own schedule skips most hourly runs; that schedule stays as a backup. Each run does the health check for every project, and from Monday at 08:00 in each client’s timezone, that client’s report. Work counts as due until it is done, so a late hour delays a report but never loses it, and every run is recorded so we can see each hour that ran. One small app on Vercel serves a dashboard with every project, the status of every source, every report, its settings and a full log. The memory is the Neon Postgres database. Nothing has to listen around the clock.

Where we are

The agent reports on InstantSchema every week and on our own site, now the first project in its launch phase, now including their email, live in their Slack channels, and the InstantSchema team follows the same numbers at app.3day.studio. In its first week it showed that new trials stopped the same day a new signup flow shipped. Next we open it to a few more companies at a time, each reviewed by hand before its first report goes out. We will write up what it found, including what it got wrong.

Get the agent for your website

We are opening it to a few companies at a time. Leave your details and we will email you when it is your turn.

Monthly visitors (optional)

No access is given until we email you.

Questions founders ask

01

What is an AI agent for website optimization?

Software that reads a website's data on a schedule, finds what limits growth and proposes specific fixes. Ours connects Google Ads, Google Search Console, PostHog, Stripe, the site's code and the project's email, posts a short report in Slack every week or two, and a person makes every change it proposes.

02

Is the agent autonomous?

It runs on its own schedule and does the analysis without anyone asking, but it never changes a site or an Ads account by itself. Every improvement has an ID, and a person makes the change in a working session with the agent's tools, and the previous state is noted so the change can be undone. Crucial problems go to a person first, who decides what the client is told.

03

Can an AI agent optimize Google Ads?

It finds searches that cost money and never lead anywhere, the keywords that bring the most conversions, and places where Ads and analytics disagree, and proposes fixes such as blocking search terms. A person makes every change, because a wrong Ads change costs money the same day.

04

How is this different from an AI website audit tool?

An audit looks at one source once. The agent joins sources every week: it can see that a paid search lands on a page that does not convert, or that Ads counts conversions PostHog never sees. It also lists what changed on the site that week and remembers what it proposed and whether it was done. Measuring each change's result 28 days later is next.

05

Does the AI write the numbers in the report?

No. Code pulls the data and computes every number, a second check re-reads the key numbers (Ads spend, clicks from Google, revenue from Stripe, and emails delivered or, for a list tool, the subscriber totals) another way, and the language model only ranks findings, explains them and drafts the text.

06

What does it cost to run?

Hosting is close to free at this scale: scheduled jobs on GitHub Actions, one small app on Vercel and a Postgres database. The real cost is the language model, well under a dollar per report, capped per run and by a monthly limit per organization.

Sources (10)

  1. Google Ads API: access levelsExplorer, Basic and Standard access, and what each allows
  2. Google Ads API: developer token sunsetAccess now lives on the Google Cloud project
  3. Google Search Console APISearch Analytics queries for clicks, impressions and positions
  4. PostHog: query APIHogQL and funnel queries over a project's events
  5. Stripe: restricted API keysKeys that can only read the resources you choose
  6. Stripe: Billing analyticsHow Stripe defines MRR, which the report follows
  7. Resend: email metricsDelivered, bounced, complained, clicked and opened counts per period
  8. Gmail API: domain-wide delegationHow a Workspace admin lets a service account read one scope
  9. Slack: chat.postMessageHow the agent posts reports and improvements
  10. NeonServerless Postgres, the agent's memory, hosted in Frankfurt
AIAI agentwebsite optimizationagentic SEOGoogle AdsSearch ConsolePostHogStripeemailSlackClaude
Summary
Share
Portrait of Jaro

Jaro

CreativeDesign9 yearsYouTube

Founder and creative director. Nine years and 60+ websites as a Webflow Premium Partner, including the rebrand and site for MobiWire, a top five ODM globally. Runs several SaaS products of his own, including a top three SEO app for Webflow.

Keep reading

Get a free website roadmap in 3 days

  • Tech stack
  • Timeline
  • Migration
  • SEO
  • Budget

Tell us about your company on a 15-minute call. Three days later you get a free roadmap showing how we would approach your website, brand and ads, based on 10 years and 60+ projects. It's yours to keep, so build it with us or take it anywhere.

  1. 01now
    Pick a time
  2. 0215 min
    Discovery call
  3. 03in 3 days
    Your roadmap
Roadmap preview