25 AI Marketing Agent Prompts That Produce Genuinely Usable Output

A marketing agent returns vague output for one reason: the instruction did not contain a stopping condition. "Analyse our SEO" has no finish line, so the agent produces a survey. "Name the five URLs that lost the most clicks in June and the most likely cause of each" has one, so it produces an answer.

Everything below follows from that. These 25 prompts are written for agents with real tool access - Search Console, GA4, SERP data, a crawler - not for chat assistants. Each one names a scope, a constraint and an output shape, because those three things are what turn a request into something an agent can finish and you can check.

Replace bracketed values with your own. Where a prompt references a date range, be explicit: relative phrases like "last month" are one of the most common causes of a confidently wrong analysis.

What makes a good AI marketing agent prompt

Four components. Miss any one and quality drops noticeably.

1. A bounded goal. Something the agent can declare finished. "Find X" beats "look into X".

2. The scope. Which property, which section, which date range, which market. Stated absolutely, not relatively.

3. The output shape. A ranked list of five, a table with named columns, a one-paragraph verdict. Without this the agent picks, and it usually picks prose.

4. The evidence rule. Tell it to cite the number behind each claim and to say "not available" rather than estimate. This single line removes most fabrication.

What you should generally not do is name the tools. A well-built agent selects tools from the goal; dictating the sequence converts it into a workflow and throws away the thing you are paying for. Anthropic's Building Effective AI Agents draws exactly this line: workflows follow "predefined code paths", agents "dynamically direct their own processes and tool usage". Prompt for the outcome, let it route.

Prompts for diagnosing a traffic drop

The highest-value category, because the correct second step genuinely depends on what the first step returned. When you are diagnosing a traffic drop, the whole point is that you do not know in advance whether the answer lives in impressions, position, click-through rate or a SERP layout change - so the prompt should name the decision you want made, not the checks you want run.

One rule governs this entire section: give absolute dates. "Traffic dropped last month" is the single most common cause of a wrong agent conclusion, because the agent's boundary and your boundary silently differ by a week and every downstream comparison inherits the error.

Organic clicks to [domain] fell between [date range A] and [date range B].
Identify the 5 URLs responsible for the largest share of the loss.
For each, state whether the cause is impression loss, position loss or CTR loss,
give the numbers behind that judgement, and rank your top hypothesis.
If the data does not support a conclusion for a URL, say so.

Compare [domain] Search Console performance for [month] this year against
the same month last year. Separate branded from non-branded queries.
Tell me which of the two moved, by how much, and whether the change is
concentrated in a small number of URLs or spread across the site.

Conversion rate on [landing page URL] dropped from [X] to [Y] between
[date A] and [date B]. Check whether traffic composition changed:
source, medium, device, country and landing-page variant.
Report only the dimensions that changed by more than 10 percent,
with before and after figures.

Audit [URL] and list every technical issue that could plausibly affect
its ranking, ordered by likely impact rather than by severity label.
For each, state what you observed, why it matters, and the fix.
Exclude anything you cannot verify from the page itself.

[domain] lost visibility for [keyword cluster]. Determine whether
competitors gained the positions we lost or whether the SERP itself changed -
new features, different intent, different result types.
Show the current top 10 and what changed.

Keyword and content research prompts

Find keyword opportunities for [domain] that we could realistically rank
for within two quarters. Compare our ranked keywords against [competitor 1]
and [competitor 2], filter out anything above difficulty [N], and cluster
the remainder by search intent. Return a table: cluster, example queries,
combined monthly volume, average difficulty, and why we can win it.

Map the query fan-out for the topic [topic]: list the sub-questions a
detailed prompt on this subject would decompose into. For each, check
whether the current top 3 results answer it directly.
Return a table of sub-question, answered by whom, and gap yes/no.

Identify keyword cannibalisation on [domain]: queries where more than one
of our URLs has received impressions in the last 3 months.
For each, name the URLs, the impression split, and which one should own
the query based on current performance.

Find striking-distance queries for [domain]: average position between
8 and 20, at least [N] impressions in the last 28 days.
Sort by impressions multiplied by the CTR gain we would expect
from reaching position 5. Show your working.

Build a content brief for [target query]. Research what the current top 5
results cover, name what all of them omit, and structure the brief around
that gap. Include the sub-questions the article must answer and the
sources worth citing. Do not write the article.

GA4 analysis prompts

Two constraints matter here. GA4 sampling and thresholding can quietly change what you are looking at, and the Data API meters standard properties at 200,000 core tokens per property per day with complex requests costing more - so a prompt that asks for six dimensions across eighteen months is an expensive prompt.

Using GA4 property [ID], compare organic search sessions and conversions
for [date range A] against [date range B], broken down by landing page.
Return the 10 landing pages with the largest absolute conversion change.
Flag any row where the session count is low enough that the
conversion-rate change may be noise.

In GA4 property [ID], identify which acquisition channels improved and
which declined over [date range], measured by conversions rather than sessions.
State the conversion event you used. If more than one conversion event
is configured, list them and ask me which to use before proceeding.

Analyse the path users take before converting on [goal] in GA4 property [ID]
over [date range]. Identify the two largest drop-off points by percentage
and give the raw counts alongside the percentages.

For GA4 property [ID], compare mobile against desktop over [date range]
for sessions, engagement rate, conversion rate and average revenue.
Tell me whether the mobile gap is widening or narrowing versus the
previous equivalent period.

Cross-reference GA4 property [ID] with Search Console for [domain]
over [date range]. Find pages that gained organic clicks but lost
conversions. For each, give clicks before and after, conversions before
and after, and one hypothesis.

Search Console prompts

Two hard limits shape every prompt in this section: the Search Analytics API returns 25,000 rows per request and retains 16 months of history. Ask for a two-year comparison and a competent agent will tell you the data does not exist. An incompetent one will estimate.

For [domain], list queries from the last 3 months with more than
[N] impressions and a CTR more than 40 percent below the average CTR
for their position band. Return query, position, impressions, actual CTR,
expected CTR, and the ranking URL.

Identify pages on [domain] where Google is ranking a different URL than
the one we intended for the query. Compare the ranking URL against the
page whose primary topic matches the query. List the mismatches.

Compare Search Console performance for [domain] across the last 4 quarters.
Identify content clusters in structural decline - falling impressions
across three or more consecutive quarters - as opposed to seasonal dips.
Show the quarterly figures for each.

For [domain], report performance on Google AI surfaces over [date range]
using Search Console generative AI reporting. Compare it to overall web
search performance for the same period. Do not use any third-party
estimate of AI visibility.

Competitive and brand visibility prompts

Compare [domain] against [competitor 1] and [competitor 2] on organic
visibility for [topic area]. Return a table of shared keywords where they
outrank us, keywords only they rank for, and keywords only we rank for,
with volume and difficulty on each.

Check how [brand] is currently described when [target question] is asked
of the major AI assistants. Report what each said, which sources it cited,
and whether our own domain was among them. Quote the responses rather
than summarising them.

Analyse the last [N] articles published by [competitor] on [topic].
Identify the structural pattern they use - typical length, heading shape,
what they cite, what they consistently omit. Do not recommend copying it;
tell me what the omission is worth.

Reporting prompts

Produce the monthly organic performance summary for [domain], [month].
Structure: one-paragraph verdict, then a table of clicks, impressions,
average position and conversions with month-over-month and year-over-year
change, then the three most significant movements with a cause for each,
then next month's recommended focus. Every number must come from the
connected data. Write "not available" where it does not exist.

Turn the analysis above into a client-ready document. Keep every figure
identical. Remove internal tool names and jargon. Lead with the business
outcome rather than the metric. Maximum one page.

Every Monday at 9am, check [domain] Search Console for any query cluster
whose clicks fell more than 25 percent week over week with at least
[N] impressions. If nothing crosses the threshold, reply with a single
line saying so. If something does, investigate the cause before reporting.

Prompt anti-patterns to avoid

How to stop an agent inventing data

Three lines, added verbatim to any prompt where the numbers matter:

Every figure must come from the connected data sources.
Where a figure is unavailable, write "not available" - never estimate.
State which source each figure came from.

This works because fabrication is usually the model resolving an unstated conflict between "be helpful" and "be accurate". Making the accuracy requirement explicit resolves it the right way. It is not a guarantee, which is why the transcript matters: an agent cannot reliably tell when it is wrong, and Gartner's assessment that most agentic deployments fail on inadequate risk controls describes precisely this gap.

How specific should a prompt be

Specific about the goal, the scope and the output. Loose about the method. That balance is easy to state and easy to get wrong in both directions.

Under-specified looks like "check our Search Console for problems". The agent has no finish line, so it samples a bit of everything and returns a list you already knew. Over-specified looks like "query Search Console for the last 28 days, then filter to position 8 to 20, then join against GA4 sessions, then rank by impressions". That is a script, and if the first pull reveals the real problem is elsewhere, the agent will follow your sequence anyway and miss it.

The version that works sits between them: state the decision you want made and the evidence standard it must meet, then let the agent choose how to get there. Practitioner guidance from Search Engine Land on building agent skills lands in the same place - scope instructions tightly enough that output is repeatable, without dictating the path.

A practical test: if you can predict the exact tool sequence your prompt will trigger, you have over-specified. If you cannot predict what the answer will look like, you have under-specified.

One more thing worth knowing

No prompt makes agent-written content more eligible for AI Overviews or AI Mode. Google's guidance states there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary". Prompts in the wild that promise "AI-optimised" output are selling you formatting. Use agents to find what to write about and to check whether you covered it - that part works.

To run these against your own connected Search Console and GA4 data, see Autopilot. If you want the reasoning behind the scoping advice, what autonomous agents can and cannot run covers the limits in detail.