How to automate weekly PPC reporting with AI
A step-by-step guide for PPC leads to automating the weekly report, where the numbers come from the platform APIs or a connector and the AI writes only the words.

PPC reporting is the weekly job of pulling spend, clicks, conversions and cost per result out of Google Ads, Microsoft Advertising, Meta and LinkedIn, setting them against the week before, and explaining what changed. AI automates it well under one condition: the numbers come from the platform APIs or a connector, and the model only writes the explanation.
You know the Monday version of this. Nine browser tabs, a spreadsheet with a formula someone broke in March, and a summary you write at 9:40 for a 10:00 send. The first reply is from sales, asking why conversions fell when they can see the same deals in the CRM. You spend the rest of the morning explaining a number that will change again by Thursday.
This guide is for the head of paid media or the PPC lead who wants that morning back. It covers what to have in place first, every route for getting the data out as of November 2026 with prices checked on the vendors' own pages, seven steps that end in a report nobody has to assemble by hand, and the ways this goes wrong. If you want the report layout itself, the field list and the prompt that fills it are in our PPC report template.
Numbers are computed and words are generated
The rule that keeps an automated PPC report honest is that every number is produced by a query, a formula or a script, and every sentence is written by the model from those numbers alone. Think of it as three layers: a data layer that pulls rows from the platforms, a numbers layer that turns them into the figures you report, including every week-over-week change and every ratio, and a narrative layer that reads a small table of finished figures and explains it. The model never adds up the raw rows and never remembers last week's total from a previous conversation.
Why so strict? Because a language model will give you a figure whether or not it has worked it out. Adam Tauman Kalai and colleagues at OpenAI and Georgia Tech put it plainly in their September 2025 paper Why Language Models Hallucinate: "the training and evaluation procedures reward guessing over acknowledging uncertainty." A model that is asked to compute a cost per lead from a messy export behaves like a student who would rather guess than leave the answer blank. You do not want a confident guess in a document your CFO reads.
Imagine two teams running the same Google Ads account. Company A pastes a 2,000-row keyword export into a chat window and asks for "a summary with the key numbers." Company B runs a query that returns twelve finished figures and asks the model to explain them, with instructions to say "not in this week's data" when a fact is missing. Both get a readable paragraph. Only Company B can defend every number in it, and only Company B can rerun it next Monday and get the same arithmetic. Scalewell helps paid teams build the report this way, so Monday's numbers assemble from the team's own data and every one of them can be defended.

Last week's conversions are still moving when you report them
Google Ads can report a conversion up to 90 days after the click, and it books that conversion to the date of the click, so last week's conversion column keeps growing after you have sent the report. Google's own data discrepancies page states both facts: "Conversions can be reported up to 90 days after the click" and "Google Ads reports conversions on the ad impression date. Other reporting tools attribute them to the conversion date." Impression date or click date, Google means the day of the ad interaction, not the day of the conversion. Google's conversion lag page adds the consequence for your Monday email: lag can "cause CPA to look inflated and ROAS to look deflated." Google's guidance on the days to conversion report suggests a date range that "ends at least 30 days ago" if you want complete conversion data.
The other platforms move on their own clocks. Meta's Insights API best practices say insights "refresh every 15 minutes and do not change after 28 days of being reported," and that metrics "may continue to update for a couple of days after an ad has completed." LinkedIn's ads reporting documentation calls account, campaign and creative metrics "near real-time," with professional demographic breakdowns delayed by 12 to 24 hours and video watch time by up to 48 hours. And if your dashboard sits in Data Studio (formerly Looker Studio), Google's data freshness page says Google Ads data refreshes every 12 hours and "That rate can't be changed."
So the weekly report needs two conversion columns, and this is the habit most published templates skip. The first column is this week, marked provisional. The second is the week before, restated with the conversions that arrived since you last reported it. After a month of doing this you have a measured lag curve for your own account, and the restatement column stops surprising anyone. B2B teams with long sales cycles feel this most, because a demo request booked on a Tuesday click often lands the following week.
Pick one conversion number before you automate anything
The one decision to make before writing a single line of automation is which conversion figure the report treats as the result, because the platforms will never agree with each other or with your CRM. Tejas Khot, director of performance marketing at Pixis, gives a clearly hypothetical example in his October 2026 guide to automating PPC reporting: "Meta reports 140 purchases. Google Ads reports 85. Your store records 112 orders." His point is that all three can be correct under their own rules, and that "Their sum, 225, is not a count of unique orders."
For a B2B team our advice is to keep platform conversions as a diagnostic and treat the CRM-connected figure as the result. The platform number tells you whether the account is healthy this week. The CRM number, meetings booked or opportunities created from paid traffic, tells you whether the spend is doing its job. Report both, label both, and never let the model combine them. If your CRM figure arrives with a delay, that is one more reason for the restated column above.
Access, a Cloud project and one conversion definition come before any automation
Five things need to exist before the automation does, and none of them is technical.
- Access that will survive staff changes. Use a shared or service login for the ad accounts, and for Google Ads make sure the manager account is linked. Data Studio's Google Ads connector documentation notes that report viewers "must also have access to the manager account and all selected sub-accounts," which is where many agency dashboards quietly break.
- A Google Cloud project, if you plan to use the API or an MCP server. As of November 2026, Google's access levels page states that "Developer tokens were sunset on September 9, 2026" and that access is "now managed per Google Cloud project," with Explorer access allowing 2,880 operations a day against production accounts, Basic 15,000 and Standard unlimited.
- The conversion definition from the section above, written down with the attribution setting and lookback window you use.
- A one-page report spec: the metrics, the comparison periods, the thresholds that count as a change worth a sentence, and the people who receive it.
- A destination: an email, a Slack channel, a Google Sheet or a shared document. Pick one and keep it for at least a quarter, because the value of a weekly report is in its sameness.
Seven routes get PPC data into a table, as of November 2026
There are seven routes for getting PPC data into a table your model can read, and we recommend none of them over the others. The right route is the cheapest one that gives you the numbers in a stable shape without the model ever touching raw rows.
| Route | What it does | Verified cost, November 2026 | Fits when |
|---|---|---|---|
| Google Ads Report Editor, scheduled | Emails a saved statistics table on a schedule. Google's help page lists frequency choices such as daily or weekly and formats such as CSV and PDF. | Free | One Google Ads account, no other platforms, someone still writes the words |
| Google Ads scripts | JavaScript inside the account. AdsApp.report() runs GAQL and exportToSheet() writes to a Google Sheet, per the scripts reporting docs. Google's scripts help says a script can run "once, daily, weekly or monthly at a certain hour," and the limits page caps a run at 30 minutes. | Free | Google only, you want a Sheet the model can read, no servers to run |
| Data Studio (formerly Looker Studio) | Free dashboards. The Google Ads connector reaches "up to 50 sub-accounts per data source" from a manager account. Google Ads data refreshes every 12 hours, fixed. | Free | A live dashboard matters more than a written narrative |
| Connector services | Pull many platforms into Sheets, Excel, Data Studio, Power BI or a warehouse on a schedule. Supermetrics: Starter €52 a month (€39 billed yearly) for 2 data sources, Growth €199 (€159) for 7, with Claude, ChatGPT or Microsoft Copilot among the destinations. Windsor.ai: a free plan with 1 source, Basic $23 a month ($19 yearly) for 3 sources, Standard $118 ($99) for 7, MCP access on every plan. Coupler.io: Starter $32 a month ($24 yearly) with daily refresh, Active $132 ($99), Pro $259 ($199) with hourly refresh. | €39 to €199 or $19 to $259 a month at the tiers listed | Two or more platforms and nobody on the team writes code |
| Agency reporting platforms | Templated client reports with scheduling, white label and AI summaries. AgencyAnalytics: $20 per client per month billed annually, including AgencyAI insights and analysis and MCP access for ChatGPT and Claude. Swydo: €69 a month (€62 yearly, shown as $69 outside the EU) covering the first 10 data sources, with AI report summaries and a Swydo MCP. DashThis: from $54 a month ($44 yearly) for 3 dashboards, AI Insights with four preset views included, AI Insights PRO add-on $19 a month. Optmyzr: an AI assistant that generates and summarizes reports across Google, Microsoft, Amazon, Meta, LinkedIn Ads and GA4, with no public price on the reporting page and a 14-day trial. | $20 per client to about €69 a month entry | An agency or an in-house team reporting to several stakeholders who want a branded PDF |
| Google Ads API plus an AI client | The Google Ads API (v26.1 released 28 October 2026 per the release notes) returns GAQL results through GoogleAdsService.Search or SearchStream. Google's official MCP server runs with pipx over stdio locally or on Cloud Run, needs a Google Cloud project ID and OAuth or application default credentials, and its documentation lists its mode as "Read and partial mutate" with a "configurable read-only mode", with the read tools list_accessible_customers, search and get_resource_metadata. The repository README also lists four asset and label write tools that propose changes by default and can be switched off, so check what the version you install exposes. Claude custom connectors are available on Free, Pro, Max, Ultra, Team and Enterprise plans, with Free limited to one connector. ChatGPT and the OpenAI API connect to remote MCP servers too. | Free software, plus any Cloud Run or AI subscription costs | Someone on the team can run Python and you want the model to query the account directly |
| Scripted pipeline | Code pulls the data, computes the facts, calls a model for the narrative and emails the result on cron. AdVenture Media's Isaac Rudansky published an eight-step walkthrough in April 2026 using Claude Code, Python and the Anthropic API, hosted on a VPS he prices at about $5 to $7 a month. | Hosting plus model usage | You want full control and already run scripts |
A few points the table cannot hold. Meta's Insights API is "available as an edge off of all ad objects" at account, campaign, ad set and ad level, and asynchronous report IDs expire after 30 days. LinkedIn's adAnalytics endpoint requires the r_ads_reporting permission, returns only impressions and clicks unless you name up to 20 metrics in the fields parameter, and keeps daily data for six months before rounding older requests to months. Both matter if you are writing the pull yourself. For the Google side in more depth, including the setup steps for each MCP route, see our guide to Google Ads MCP servers.
On time saved, no independent study exists that we could find. Rudansky's walkthrough puts manual reporting at 3 to 6 hours a week against about 20 minutes automated, and Dataslayer's July Cintra, in a January 2026 comparison, estimates 3 to 5 hours a week manual against about 30 minutes with an automation platform. Both come from parties with something to sell, so treat them as the shape of the saving and measure your own.
Seven steps to a weekly report that writes itself
The steps below work with any route from the table, because each one ends in something you can look at before you build the next.
-
Write the spec and the thresholds. Decide the metrics, the two comparison periods (last week against the week before, and against a four-week average), and the size of change that deserves a sentence. Rudansky's spec, for example, flags a return on ad spend drop over 20% week over week and a click-through rate drop over 25%. You end this step with one page everybody has read.
-
Pull a fixed window. Use the platform's own definition of last week so the report never shifts with the day it runs. In GAQL that is
LAST_WEEK_MON_SUN, which Google's date ranges page defines as "The 7-day period starting with the previous Monday." A campaign-level query looks like this:
SELECT
campaign.name,
campaign.status,
metrics.impressions,
metrics.clicks,
metrics.cost_micros,
metrics.conversions,
metrics.conversions_value,
metrics.conversions_by_conversion_date
FROM campaign
WHERE segments.date DURING LAST_WEEK_MON_SUN
AND campaign.status = 'ENABLED'
metrics.cost_micros is cost in millionths of your account currency, so divide by 1,000,000 before anyone sees it. metrics.conversions_by_conversion_date gives you the second clock from the lag section.
Meta and LinkedIn have no GAQL, but each can hand you the same Monday to Sunday week. On the connector route the connector pulls all three platforms on its schedule and this step is a column rename, so move on to step 3. If you write the pull yourself, Meta's Insights API lists last_week_mon_sun among its date_preset values as of November 2026, and a time_range with since and until dates replaces the preset when you move the window back in step 6. LinkedIn's adAnalytics endpoint has no week preset, so you write the Monday and the Sunday into dateRange yourself, both dates inclusive, and name the metrics in fields or it returns impressions and clicks alone. For the week of October 26 to November 1, 2026, the two calls look like this:
# Meta, campaign level, last full Monday to Sunday week
GET https://graph.facebook.com/v27.0/act_<AD_ACCOUNT_ID>/insights
?level=campaign
&date_preset=last_week_mon_sun
&fields=campaign_name,spend,impressions,clicks,actions,account_currency
# LinkedIn, campaign level, the same week with the dates written out
# headers: Authorization: Bearer <token with r_ads_reporting>,
# Linkedin-Version: <YYYYMM>, X-Restli-Protocol-Version: 2.0.0
GET https://api.linkedin.com/rest/adAnalytics
?q=analytics
&pivot=CAMPAIGN
&timeGranularity=ALL
&dateRange=(start:(year:2026,month:10,day:26),end:(year:2026,month:11,day:1))
&accounts=List(urn:li:sponsoredAccount:<ACCOUNT_ID>)
&fields=pivotValues,impressions,clicks,costInLocalCurrency,externalWebsiteConversions
Meta's spend and LinkedIn's costInLocalCurrency both arrive in the account's currency, so only Google needs the division. Without the APIs, Meta's Ads Reporting exports a saved report as an .xlsx or .csv file and can email a single-account report on a schedule to people with access to that account, and LinkedIn's Campaign Manager exports a CSV from the Advertise tab once you set the time range, pick a report type and choose a Daily, Monthly or All breakdown. You end this step with one table per platform, all using the same column names.
-
Compute the facts in code or formulas. Spend, conversions, cost per conversion, the week-over-week change for each, the pacing against the monthly budget, and a flag column that compares each change with the thresholds from step 1. Add a short list of what was changed in the account last week, pulled from the change history or typed by whoever made the changes. You end this step with a facts table of perhaps 15 rows and a changes list. Keep every week's table, because step 6 needs last week's.
-
Reconcile once against the platform interface. Open Google Ads, set the same date range, and compare spend, clicks and conversions with your table. Rudansky's instruction is the right one: "Confirm your data pulls are accurate before building the narrative layer." Time zone and attribution settings are the usual culprits when they differ. You end this step with a tolerance you have written down, and a pass.
-
Let the model write from the facts table only. Paste or pipe the facts and the changes list into the prompt below, which is deliberately boring.
You are writing the commentary for a weekly paid media report.
Rules:
- Use only the numbers in the FACTS table. Do not calculate, round, estimate or recall any number that is not in it.
- If a figure you need is missing, write "not in this week's data" instead of a number.
- The current week's conversions are provisional. Say so once.
- For every row where FLAG is true, give the most likely cause from the CHANGES list and say what we will do next week. If no change explains it, say "cause unclear, investigating".
- Write 150 to 250 words in plain sentences. No exclamation marks, no superlatives, no advice beyond the next-week actions.
FACTS:
<paste the facts table>
CHANGES:
<paste the list of account changes made last week>
You end this step with a paragraph you can check line by line against the table in under a minute.
-
Add the restated column. Rerun step 2 with the date range moved back one week, compare the result with the facts table you kept seven days ago, and show both figures. You end this step with a report that admits its own lag instead of being caught out by it.
-
Schedule it, then keep a human on it for a month. Use the scheduler your route gives you: the Google Ads scripts scheduler, the connector's refresh, the agency platform's send, or cron. Read every report before it goes out for four weeks and log how long that takes. Pacing alerts that need to fire between reports belong in a separate check, which we cover in automating PPC budget pacing. You end this step with a weekly send and a number for the minutes it now costs you.
The seven ways automated PPC reports go wrong
The most common failure is letting the model touch the raw rows, and it is also the quietest, because the paragraph it produces reads fine. The Kalai paper explains why the figure arrives with confidence either way. Keep the arithmetic in the numbers layer and this problem disappears.
Adding platform conversions together is the second. The Pixis example above exists because people do this every week, and the sum reads as a result when it is two overlapping claims on the same orders.
Comparing this week with last week without restating last week is the third. Google's conversion lag means the current week will almost always look worse than it is, and a model asked to explain a drop will find a cause for a drop that is not there.
Refresh cycles are the fourth. A Data Studio page refreshes Google Ads data every 12 hours, so a report pulled early Monday can be missing part of Sunday, and Meta's figures can keep moving for a couple of days after an ad finishes. Pull the weekly window on Tuesday if you can live with it, or label Monday's version provisional.
History that changes under you is the fifth. Google's discrepancies page warns that "Switching between attribution models can affect your conversion data," so a change to attribution or to a conversion action's settings will quietly rewrite earlier weeks. Put those changes in the changes list and the model has a cause to point at.
Thresholds set so tight that every week is an alert are the sixth, because they train your readers to skip the commentary. Start loose and tighten once you know the account's normal swing.
And the seventh is the report that shows 400 leads and says nothing about what sales did with them. If the CRM-connected figure is missing, the automation has saved you time producing a document that answers no one's question.
The automation worked if the totals match for four weeks running
A weekly PPC report that is automated well passes four checks, and none of them is "it looks good." First, the platform totals in the report match the platform interface within your written tolerance for four consecutive weeks. Second, the restated column moves by roughly the same amount each week, which means you understand your lag. Third, the minutes you log in step 7 fall and stay down. And fourth, the questions you get back change from "is this number right?" to "what are we doing about campaign X?", which is the only kind of question a report should generate.
If you want to put a value on those minutes for a budget conversation, our piece on measuring the ROI of AI in a paid team walks through how to do it without inventing a number.
A paid Google Ads reporting tool is worth it once you report on more than one platform
A dedicated Google Ads reporting tool pays for itself at the point where a second platform or a second stakeholder enters the picture, because the free routes above cover one account well and several accounts badly. For a single Google Ads account, Report Editor schedules, a script writing to a Sheet, and a free Data Studio page cover the data layer at no cost, and a model subscription you already pay for covers the words.
Imagine a team that spends four hours a week assembling a Google, Meta and LinkedIn report by hand. At a fully loaded $60 an hour that is about $960 a month, which is the comparison to hold against a €52 connector seat or a $20 per client agency platform. The hypothetical is ours and the hours are yours to measure.
How Scalewell gets the Monday report writing itself
Scalewell helps paid teams turn the weekly report into something AI drafts before the Monday meeting, so the morning becomes a quick read and a send. The audit maps your team's recurring jobs and scores where AI should start. When reporting ranks first, the build sprint turns it into a workflow on your own ad accounts, with the numbers computed and the model writing the words.
Your paid media team checks each report against its own figures until the totals match. The engagement runs typically 8 to 12 weeks, and afterwards the team can change what the report covers without help.
Where to start
Start by writing down the conversion definition and the one-page spec, then pick the cheapest route in the table that gets the data into a table without the model touching it. Scalewell sets this kind of reporting method up for paid media teams, built around how the team already works.
Frequently asked questions
How often should PPC reports be sent?
Send a short automated report weekly and a fuller one monthly. The weekly version catches pacing or tracking problems within days, while the monthly version has enough settled conversion data to judge results, since Google Ads can report conversions up to 90 days after the click. Daily reporting suits launches and tests, and belongs in an alert.
What should a weekly PPC report include?
A weekly PPC report should include spend against budget pacing, clicks, platform conversions marked provisional, the restated conversions for the previous week, cost per conversion, the CRM-connected result such as meetings or opportunities, the week-over-week change for each, and a short commentary that names the account changes behind any flagged movement and the action planned for next week.
Can Claude or ChatGPT read my Google Ads account directly?
Yes, as of November 2026, through an MCP server. Google publishes an open-source Google Ads MCP server. Its documentation lists its mode as "Read and partial mutate" with a "configurable read-only mode", and the current repository ships asset and label write tools that propose changes by default and can be disabled, so check your installed version. Claude and ChatGPT connect to remote MCP servers, as do several connector vendors.
Is Data Studio (formerly Looker Studio) enough as a Google Ads reporting tool?
Data Studio, which Google renamed from Looker Studio in April 2026, is enough for a free single-platform dashboard. Its Google Ads connector reaches up to 50 sub-accounts per data source, viewers need access to those accounts, and Google Ads data refreshes every 12 hours with no option to change the rate. It writes no commentary, so pair it with a script or connector that produces the facts table.
How do I check that an AI-written PPC report is accurate?
Check every number in the commentary against the facts table it was given, which takes under a minute. Then compare the table's totals with the ad platform interface for the same date range and attribution setting. Any number in the text that is missing from the table was calculated or guessed, so tighten the prompt.