AI Doesn’t Lack Intelligence—It Lacks These 11 Growth Data Sources

Give AI only a screenshot of an ad dashboard and ask it to analyze campaign performance, and the result usually won’t be very good. In a tweet, Will Hu puts the problem bluntly: the model doesn’t lack intelligence; it lacks data. The screenshot contains only spend, clicks, and the conversion numbers the platform provides. Where users came from, what they viewed, which day they stayed until, and whether they ultimately paid are all blind spots. AI can only follow that screenshot and invent a conclusion that looks reasonable.

Will Hu’s list contains 11 data sources. 1 through 6 are your own; 7 and 8 determine whether the data is accurate; 9 and 10 are other people’s; and 11 is how to slice it. Whatever piece is missing, the judgment will be distorted in that direction.

1. Website behavior: Google Analytics or Plausible. Look at sitewide traffic—where people come from, what they view, and at which step they drop off. You can cross-slice by channel, region, device, page, and language.

2. Search performance: Google Search Console or Bing Search Console. Look at organic search—which terms users search to find you, and which pages get impressions but no clicks. You can likewise cross-view by query, page, country, and device.

3. App attribution: Adjust or AppsFlyer. See where mobile installs come from. If you run an app without an MMP, the model doesn’t even know which channel or campaign brought the user.

4. In-product behavior: Mixpanel, Amplitude, or PostHog. See what features people use after they arrive and which day they are still around. Being able to bring people in does not mean you can make them stay.

5. Ad platforms: Meta Ads, Google Ads, TikTok Ads. Look at the money spent; views, clicks, and spend can be broken down by campaign, ad set, and individual creative. This is the dataset most teams are least short of, and it is also the one most easily mistaken for the whole picture.

6. Payment truth: Stripe, Paddle, PayPal. Look at the money collected—who actually paid and how much. For ecommerce, look at Shopify and Amazon; WordPress and Wix sites also have their own order data.

7. Postback channel: Pixel to API or S2S. It determines what the ad system can learn. Beginners should set up the Pixel first; once mature, move to server-side postbacks. Meta calls it the Conversions API, and TikTok calls it the Events API.

8. Attribution definitions: view or click, and what attribution window in days. Platforms each claim credit for themselves. If the definitions are not unified, the model will add several platforms’ numbers together and reach the wrong conclusion.

9. Third-party traffic and search: Similarweb, Ahrefs, Semrush. Look at competitors’ traffic scale, channel mix, backlinks, organic search terms, keywords, and search ads. These are all estimates; they are suitable for seeing trends and comparing relative size, not for use as a ledger.

10. Competitor creatives: Meta Ad Library, Google Ads Transparency Center, TikTok Creative Center. They are free and publicly available from the platforms themselves, and you can see what creatives others are running and what selling points they are emphasizing.

11. Breakdown dimensions: device × country × selling point × objective × SKU × category. Cross all the preceding data into one table. Categories can be further split by industry, competitor, and scenario. Each cell ultimately lands on two numbers: number of paying users and payment amount.

The comments added a few common pitfalls. FlowOps Daily reminded that the same order may be claimed by several platforms at once. First figure out whether each platform counts views or clicks and what attribution window it uses, and only then let AI analyze. The neater the table looks, the harder duplicate counting is to spot.

Citeshine added AI search: in GA4, filter session source for https://chatgpt.com and look at landing pages. OpenAI’s publisher FAQ says that referral links from ChatGPT automatically carry a utm_source, so they can be separated out. Looking only at clicks misses a piece: if a user reads the answer and doesn’t click through, neither GA nor GSC can see it, so you have to ask the buyer’s questions yourself.

Albert mentioned that attribution is often ignored. Without attribution, you cannot calculate the value and payback of each channel. When only running ASA, Apple’s built-in attribution API is free and accurate enough; only multi-channel campaigns need Adjust. Adjust’s official annual fee starts at $18,000, with pricing ranging from $0.026 to $0.03. Smaller teams can consider splitting the cost with a team that has already purchased it.

JokerMaycry suggested building your own knowledge base as well: a detailed product description in an MD document that spells out features, limitations, and scope of application; complete transaction information—for consumer products, first-touch and last-touch attribution; for B2B, retained inquiries that form orders; for SaaS, first visit and feature usage before registration or purchase; for SEO growth, cross-analyze GSC data with real website logs. Based on these, ad automation can be targeted.

The list itself is not complicated; what’s hard is getting 7 and 8 right. The postback channel determines what the ad system can learn, and attribution definitions determine who gets credit. If the definitions are not unified, the more complete the data in 1 through 6 is, the more neatly the errors line up. Third-party data is fine for seeing trends; don’t treat estimates as a ledger. Once all the pieces are in place, AI has a chance to make judgments; reading reports is only the shallowest layer. With one fewer dataset, it has one more reason to guess.