Custom AI Agents for Marketing

Q: Can these agents produce inaccurate data?

Guardrails are written into the tooling. The agents are instructed never to output invented search volume, keyword difficulty, or competitor ranking claims, and never to guess at URLs the client did not supply. Output still gets reviewed before it reaches a client deliverable.

Q: What does an agent build cost?

Agent projects get quoted after a diagnosis, because cost depends on the data sources involved and the size of the site. Fixed-price audit packages include the site-wide page scoring crawl and its Excel export.

An agent that runs on a schedule is only useful if you can trust what it produces without checking every run. Ivan Sp writes AI agents that perform marketing work on a schedule instead of on request. Current builds cover keyword research pipelines, content brief generators, rank monitoring bots, internal link discovery tools, content gap automation, and site-wide page scoring. Each one runs against live data through the Google Search Console API.

What a Marketing AI Agent Actually Is

An AI agent is a program that collects data, applies a language model to it, and produces a structured output. No person drives each step. The distinction that matters is the data source. A chatbot answers from what it was trained on. An agent queries the Google Search Console API, crawls the live site with Screaming Frog, or reads a BigQuery table. Each run reasons over what it found today.

The stack behind these builds is PHP, MySQL, Python, BigQuery, the OpenAI API, and Google’s own APIs. A programming background is what makes the difference here. Most SEO consultants can describe the workflow they want, and cannot build the software that runs it.

The Agent Library

Six agent types make up the current library, each replacing a task that otherwise consumes hours every week:

Agent	What it does
Keyword research pipeline	Collects query data at volume, applies a curated stop-word list so filler terms do not dominate frequency analysis, and groups queries by search intent.
Content brief generator	Produces a brief covering target keyword, search intent, reading level, heading plan, and the specific gaps a page must close against what already ranks.
Rank monitoring bot	Tracks position changes continuously and flags movement rather than producing a report nobody reads.
Internal link discovery	Finds pages that should link to each other based on topical relationship, and identifies orphaned pages carrying no internal links at all.
Content gap automation	Compares existing coverage against demand and against competitor URL sets, marking which topics are covered, thin, or missing.
Site-wide page scoring	Crawls every page and scores it, then exports the full set to Excel for review.

AI-Scored Site Crawls Across 50 to 2,000 Pages

The page scoring agent is the one clients see first, because it ships inside every audit engagement. Every page on the site gets crawled and scored for NLP friendliness, helpfulness, originality, potential legal issues, and spam signals. The full set exports to a custom Excel file the client keeps.

The same pipeline runs parallel specialist passes across sites from 50 to 2,000 pages, with each pass asking a different question of the same crawl. The search pass classifies query intent per page and suggests schema markup. The content pass scores depth and detects decay. The paid media pass scores landing page message match. The conversion pass audits cognitive load and trust signals. Reviewing 2,000 pages by hand across four disciplines is not feasible, so the agent triages first and human review concentrates on flagged pages.

The Guardrails Are Part of the Build

Every agent carries explicit instructions never to invent data, and those instructions exist because the failure mode is silent. An agent that fabricates a search volume figure produces a deliverable that looks correct and misleads a real business decision. Three rules are written into the tooling:

Never output invented search volume, keyword difficulty, or claims about what a competitor ranks for
In audit mode, never guess at URLs the client did not actually provide
Treat client-supplied files as authoritative, and verify anything not covered by them before writing it

Competitive analysis follows the same discipline. A competitor coverage comparison maps only against a sitemap or URL list the client pastes in, which keeps it fully verifiable. The comparison never estimates competitor traffic or rankings. That restraint costs a more impressive-looking deliverable and produces one that holds up when a client checks it.

Monitoring Agents That Catch Change Early

Two monitoring builds run on a weekly cycle rather than on demand. A SERP fingerprinting agent records the ten ranking URLs for a target query. The agent also maps which search features appear, including featured snippets and People Also Ask blocks. Re-running weekly surfaces new competitors and format shifts as they happen. Waiting for a monthly report shows the same change only after a ranking drop.

A cannibalization dashboard tracks which pages compete for the same target phrases over time. On one e-commerce catalogue, several in-ear monitor product pages were competing for identical terms, which became the first item on that client’s roadmap. Cannibalized pages split ranking signal, so both pages usually underperform against what a single consolidated page would achieve.

Agents Built for In-House Teams

Marketing teams with their own headcount often need tooling more than they need another consultant. Agent builds get delivered as working software the team runs itself. The prompts, scoring rubrics, and guardrails arrive documented, so an internal developer can extend them. One recurring build converts long-form industry video into a scannable database. The agent reads a transcript, extracts every actionable tactic, tags it by difficulty and category, and exports the set as CSV.

Advisory engagements pair the tooling with direction, where Ivan Sp trains the team and reviews output rather than performing the work. That arrangement suits companies that have people and lack a senior search specialist to point them.

What These Agents Do Not Do

An agent automates judgment-free work, not judgment. Topic selection, positioning, and the decision about which pages deserve to exist stay with a person. Those calls depend on business context no crawl contains. Agent output is a first pass that concentrates human attention. Treating it as a finished deliverable is the most common way automated SEO work goes wrong.

These builds also do not replace the standard toolset. Screaming Frog, Google Search Console, Google Analytics 4, Ahrefs, and SEMrush stay in use. The agents fill three gaps those tools leave. Those gaps are log file analysis, content scoring at page level, and reporting pipelines that join data sources no single product connects.

Agent Builds in Detail

Each agent in the library is documented separately:

What a Marketing Agent Is Made Of

The trigger: A schedule or an event, rather than a person deciding to run something.
The data source: A live API rather than a static export, so output reflects current state.
The processing step: Filtering, clustering, or comparison, done in code rather than by a model.
The model step: Used where language judgement is genuinely needed, and nowhere else.
Guardrails: Explicit rules on what the agent may not output, checked before anything ships.
Halt conditions: The circumstances under which the agent stops and asks rather than guessing.
The output: A structured artefact a person can check, rather than a finished decision.
What is never automated: The judgement about whether a finding matters commercially.

Test Any Agent You Already Run

Ivan Sp suggests a test that exposes an unreliable agent in about ten minutes.

Run your agent twice on identical input, an hour apart.
Compare the two outputs line by line.
Pick three figures from the output and try to trace each to its source.
Note any figure you cannot trace.

Different output from identical input means the deterministic work is being done by a model. An untraceable figure is one you should never have acted on.

What This Looks Like on a Monday

Consider an in-house SEO who built an agent to draft weekly reports. Every Monday she reads the output and checks each figure against the source anyway.

She automated it to get her Mondays back. Verifying every run costs more than writing the report did, which is the opposite of what she was promised.

I move the counting, filtering, and clustering into code and leaves the model the writing. The same input now produces the same output, and every figure traces to a source.

Where the Standard Automation Advice Breaks Down

Would you act on your agent’s output without checking the numbers?

The usual advice: Give the model the task and the data, and let it work out the rest.
What holds up in practice: A model asked to do everything will invent the parts it cannot verify. Reliable agents do the deterministic work in code. The model handles only the judgements that genuinely need language.

Picture an agent asked to research keywords, cluster them, and write briefs, all in one prompt. The first output looks impressive.

The clustering is inconsistent between runs, and two of the search volumes were invented. Nothing in the output indicates which parts were real.

I pull the data through an API, clusters it in code, and use the model only for the language step. Every figure in the output then traces to a source.

An agent that cannot be trusted has to be checked every run, which costs more than doing the work by hand. Reliability is the whole value proposition.

An automated report nobody trusts is a slower version of doing it manually, with the added risk that somebody eventually acts on it anyway.

Four Things Commonly Assumed About Marketing Agents

Which of these describes automation you have tried?

I meet these four assumptions on most automation conversations.

Commonly assumed	What actually holds
The model should do the whole task.	Deterministic steps belong in code. A model asked to count, filter, or cluster will produce different answers on different runs.
More capable models remove the need for guardrails.	Guardrails constrain what may be output regardless of capability. A capable model still needs telling what it must never invent.
Automation replaces the analyst.	Agents remove collection and formatting work. Deciding which finding matters commercially stays with a person.
An agent is built once and left running.	APIs change, rate limits bite, and sources go stale. An agent needs monitoring as much as any other system.
Output that looks right is right.	Plausible output is the failure mode rather than the success case. Traceability to a source is what makes output checkable.

An agent either saves you a day a week or costs you one. The variable is whether you have to verify every run.

A Worked Example: Two Ways to Build One Agent

Consider a keyword research agent built two ways. The comparison below is illustrative, written to show where reliability comes from. No client engagement is described here.

Step	Model does everything	Code plus model
Data collection	Model recalls or invents figures	Search Console API, live
Filtering branded terms	Model judges each term	Regular expression in code
Clustering	Model groups by similarity	Ranking overlap, computed in code
Writing the summary	Model	Model, from the computed output
Same input, same output	No	Yes
Every figure traceable to a source	No	Yes
Needs checking every run	Yes	No

I use the model for the final row of work rather than the first. The last two rows are what make an agent worth running unattended.

Asking a model to do arithmetic is asking a novelist to audit your accounts. The prose will be excellent and the numbers will not reconcile.

The clustering row is the one that surprises people. Grouping by ranking overlap is a computation rather than a judgement, and code does it identically every time.

How an Agent Gets Built

I separate the deterministic steps from the language steps before writing anything.

Write down what the agent must produce, and what a person will do with the output.
Split the workflow into steps that are deterministic and steps that need language judgement.
Implement every deterministic step in code, so the same input produces the same output.
Connect live APIs rather than static exports, so output reflects current state.
Handle rate limits and retries explicitly, since a silent failure produces a partial report.
Use the model only where language judgement is genuinely required.
Write guardrails naming what the agent must never output, including invented figures.
Define halt conditions, so the agent stops and asks rather than guessing.
Make every figure in the output traceable to the source it came from.
Run the agent alongside the manual process for a period before trusting it unattended.

How This Engagement Is Scoped

I scope agent builds by the number of deterministic steps rather than by model complexity.

What arrives: A working agent, its guardrails and halt conditions in writing, and a traceable output format.
Typical duration: Four to eight weeks per agent, including a parallel running period.
What the client supplies: API access, the questions the agent should answer, and review of early output.
Included: Workflow analysis, code implementation, model integration, guardrails, and monitoring.
Not included: Replacing analyst judgement, and any output the agent cannot trace to a source.
What changes the estimate: Several data sources, strict rate limits, or a workflow with many judgement steps.
How it is quoted: After a short scoping call, once the workflow is mapped.

Tools and Primary Sources

The references below cover the interfaces these agents run against. Each opens in a new tab.

Search Console API returns query and page data at volumes the interface cannot display.
GA4 Data API supplies conversion and behaviour data programmatically.
OpenAI API documentation documents the model interface, including structured output and function calling.
BigQuery is where joined data lands when reporting exceeds interface limits.
Google Search Central documentation is the reference any agent output about search behaviour is checked against.

Common Questions About Custom AI Agents

Who owns the agent after the project ends?

The client owns the build. Agents get delivered as working software, with the prompts, scoring rubrics, and guardrails documented. An internal developer can run and extend them without further involvement.

How is an AI agent different from using ChatGPT directly?

A chatbot answers from training data and whatever gets pasted into it. An agent connects to live sources such as the Google Search Console API, a Screaming Frog crawl, or a BigQuery table. Every run applies the same reasoning steps. That design produces repeatable output against current data rather than a different answer each time.

Can these agents produce inaccurate data?

Any language model can generate confident text that is wrong. The guardrails are therefore written into the tooling rather than left to the operator. The agents are instructed never to output invented search volume, keyword difficulty, or competitor ranking claims. They also never guess at URLs the client did not supply. Output still gets reviewed before it reaches a client deliverable.

What does an agent build cost?

Agent projects get quoted after a diagnosis. Cost depends on the data sources involved, the size of the site, and whether the client needs a one-time build or ongoing operation. Fixed-price audit packages include the site-wide page scoring crawl and its Excel export.

Do these agents replace an SEO analyst?

No. Agents remove collection, filtering, and formatting work, which is most of the hours and none of the judgement. Deciding which finding matters commercially stays with a person.

Discuss an AI Agent Build

The first call is a free 30-minute consultation covering which parts of your current marketing work are genuinely repeatable and which need a person. Not every manual task is worth automating, and that assessment comes before any quote.

Book a free 30-minute consultation

Agent output feeds directly into the page generation described under programmatic SEO systems built at scale. Structuring the resulting content for citation is covered under AI search optimization for answer engines.