Reading Google and Meta together

The connection is the easy half

Getting an assistant to reach both ad platforms used to be the whole job. That part is now close to solved: there are MCP servers, exports, connectors and consoles, and the options are laid out here. What nobody ships with the connection is the set of rules for reading two accounts that were never designed to be read together.

So the failure mode changed. An assistant with one platform gave you a partial answer and you knew it was partial. An assistant with both gives you a complete-looking answer, in one paragraph, with a total at the end. The total is the part to distrust.

Every mistake below has the same shape. Each platform is answering a question about its own ads only, and the assistant is treating the two answers as if they came from one system.

Three words this page uses precisely

Declared

The platforms a record says the account runs. Somebody typed it in once, and nothing checks it again.

Connected

The platforms whose permissions exist right now and would answer if you called them. When the two lists disagree, this is the one that is true.

Collection gap

A date with no rows because the import never ran, as opposed to a date with rows that are all zeros. On a chart the two are the same dip.

What survives being added, and what never does

Google Ads and Meta each report on their own ads only, and neither platform can see the other. That makes some of their numbers addable and others not, and the dividing line is whether two platforms can both claim the same event. Cost is safe, because each platform bills its own spend and no euro can be counted twice by two systems. Clicks are safe too, since a click happens on one platform or the other, and impressions add up arithmetically even if reading them as reach still misleads. Conversions are not safe, because both platforms are willing to take credit for the same sale, and revenue, ROAS, CPA and CPL all inherit that double count. One line covers the whole table: add what a platform spent or served, and never add what it claims it caused.

The table has no middle column for judgement. A metric is either free of double claims or it is not, and no amount of care with the numbers moves a row from one side to the other.

Mistake one: adding conversions

Each platform decides which of its own touchpoints deserves the credit, and each uses a window it was given separately. Google documents its own default plainly:

“If you don’t customize the click-through conversion window when you create a new conversion, the default window is 30 days.” And: “A March 13 conversion from this interaction, which wasn’t counted within the last window, won’t be retroactively counted.”

Google Ads Help, About conversion windows (retrieved 28 August 2026)

Meta carries its own attribution setting, chosen in its own interface, and documented on its own help pages. The exact numbers matter less than the structure, and the structure is the problem: two windows, set independently, over one buyer who does not know either exists.

The fix is not a better model or a longer window. The fix is a rule the assistant is told before it answers: cost, clicks and impressions add; conversions, revenue and every ratio built on them do not.

Mistake two: the declared list beats the connected one

Every account record somewhere carries a field like active platforms. Somebody typed it in when the account was set up. The connections are a different thing entirely: the permissions that exist right now, the ones that would answer if you called them.

Those two lists are supposed to agree. In our own build they have disagreed, and the cost was specific: a report was written from the declared list and left out 27 % of the spend. The report looked complete. Nothing in it said a platform was missing, because as far as the record was concerned nothing was.

Why the assistant picks the wrong one

A declared list is a tidy field with a tidy answer, so it reads like the authoritative one. The connected list has to be derived, it comes back messier, and it sometimes contradicts the tidy field. An assistant asked which platforms does this client run will reach for the tidy field unless told otherwise.

Our own guide for models says it in one line, and it is the line worth stealing: what is connected wins. A platform that is declared and not connected is a platform with no data, and a report that does not say so is a report with a hole in it.

Mistake three: a gap read as a fall

A chart cannot tell you the difference between a day with no spend and a day that was never collected. Both arrive as a zero, both draw the same dip, and the second one is not a fact about the account at all.

How the two look on the way in

• A real zero comes with rows. The campaigns exist, the day exists, and every number on it is a zero.
• A collection gap comes with nothing. No rows for that date, because the import did not run or the platform was slow to close the day.
• A summary flattens both. By the time either reaches a chart they are the same pixel, and the chart cannot tell you which one it drew.

The question to make the assistant answer before it interprets any curve: how many days of this period actually have rows, and what is the status of each source? An answer that cannot name the coverage is an answer about a chart, not about an account.

Collection lag also makes the freshest day of any period the least trustworthy one. Ad platforms keep adjusting a day after it closes, and conversions in particular arrive late. The last two or three days are provisional, whatever the chart implies.

The block to paste in before you ask anything

None of the three mistakes needs a smarter model. Each one needs a rule stated in advance, because an assistant given two totals and no rules will add them. Paste this into the context, the project instructions or the system prompt, wherever your assistant keeps what it should already know.

ADD ACROSS PLATFORMS   cost, clicks, impressions

NEVER ADD              conversions, revenue, ROAS, CPA, CPL
                       Each platform attributes with its own window and
                       cannot see the other, so one sale can be claimed
                       twice. Report each platform separately instead.

BEFORE ANY NUMBER      Name the platforms that are CONNECTED, not the
                       ones a record declares active. If the two lists
                       disagree, the connected list wins, and say so.

BEFORE ANY TREND       State how many days of the period have rows. A
                       day with no rows may be a day that was never
                       collected. Say which one it was.

TREAT AS PROVISIONAL   The last 3 days of any period, and any
                       conversion metric inside them.

WHEN THEY DISAGREE     Give both figures, separately, then name one
                       business-side number as the tiebreaker:
                       sales, or qualified leads out of the CRM.

NEVER                  Invent a total that no platform reported.
                       If the honest answer is two numbers, give two.

Questions people ask before trusting a cross-platform number

Related reading

Related reading. Marketing attribution: the four dials that make one platform disagree with itself. Connect your AI to your ad data: the three ways to get an assistant to the numbers in the first place. Google Ads MCP: which servers can change your account and which only read. What not to automate: the jobs that stay yours once the reading is handled.