Connecting AI Video to Traffic and Sales: Tracking Design, Social Data Integration and Conversion-Based Optimization

AI can produce video in volume, but the sale happens after the video, on the landing page, the product page and the checkout. To work out how each video performed, the site-side events, parameters and data tables have to be in place before any automatic optimization. This article covers the architecture of that closed loop, the tracking fields, how the data is merged, the rules for allocating traffic, the two checkpoints automation must leave to people, and finally a rollout order. For keeping characters and scenes consistent on the video side, see Character and Scene Consistency in AI Video.

Video earns attention; the sale happens on the site and the product page

Creative has a large effect on results. A Nielsen, 2017 article states that creative quality contributes about as much to a brand's market results as the other factors combined: reach, target audience, timing and context. Google cites Link AI research done with Kantar: YouTube ads that follow ABCD (Attract, Brand, Connect, Direct) were predicted by the model to raise short-term sales likelihood by 30% and long-term brand contribution by 17% on average (Google and Kantar, 2021). Both figures are model predictions, not measured sales.

The video format also has to fit the product. Yang et al., 2025 analyzed sales data from 564 short videos and live streams on TikTok. Short videos suit products that can be searched and price-compared, while live streams suit products that need to be experienced, most clearly high-priced ones. Before choosing a format, decide which kind of product it is.

Each platform's built-in shopping mechanism works differently. With YouTube product tagging, long videos show a shopping button and Shorts show a product sticker, a single piece of content can tag up to 60 products, and timestamp tagging is limited to long videos over one minute (YouTube Help: Tag products). In the YouTube Shopping affiliate program, YouTube handles affiliate links, conversion tracking and billing, the merchant pays a commission on a sale, and checkout still happens on the merchant's own site (Google Merchant Center Help). Demand Gen in Google Ads puts video and image assets in one campaign served across YouTube, Discover, Gmail and Display, and product promotion assets for Shorts and pre-roll require Merchant Center product data (Google Ads Help: Demand Gen).

Check where a commerce figure comes from. In online A/B tests on its own e-commerce system, a Kuaishou research team reported that generative recommendation raised product-card GMV by 14.7%, short-video GMV by 10.3% and live-stream GMV by 4.9% (Kuaishou research team, 2026). That is the result on Kuaishou's own system, the test period and traffic share are not stated, and it cannot be applied directly to another merchant. The landing page after the video and its tracking decide whether performance can be calculated at all.

Architecture: production, distribution, site and data

The loop has four stages: content production, distribution channels, the site and product pages, and the data layer. Each handoff carries different data. Production hands distribution a video with a creative ID, distribution hands the site a link with utm parameters, the site hands the data layer e-commerce events and orders, and the data layer hands production a weekly decision to keep, rewrite or retire.

Of the four stages, only the site and product pages are fully controlled by the business, so whether the whole chain connects depends on the events and parameters there. Platform view data comes from each platform's API and site conversions come from the site's own events, and the two sides can be matched only through the same creative ID.

Tracking design: every video traceable to an order

The first step is to assign a creative ID to every video version. Every link carries utm_source, utm_medium and utm_campaign (Google Analytics Help: custom URLs), and the creative ID goes into utm_content. If one video is cut into three opening versions, there are three utm_content values, which later shows which version brought the order.

The landing page records the utm parameters, the GA4 client_id and the Google Ads GCLID, stores the three values in the session or checkout data, and sends the values along with the order when it is placed. These three values are what later connect an order back to a video. If any one is missing, server-side forwarding and offline conversion import have nothing to match against.

Events follow the GA4 e-commerce events (Google, Measure ecommerce). The purchase event requires transaction_id and items, and currency is required when value is set (Google, GA4 Events reference).

Sending events from the browser alone misses blocked requests and orders placed after returning from a third-party checkout. The fix is server-side: server-side Google Tag Manager collects and forwards user behavior in a server container (Google, Tag Manager Server-side) and can serve as a single forwarding point. The Meta Conversions API connects server-side data to Meta (Meta, Conversions API), and when it is sent together with the browser Pixel, duplicates are identified and removed by event_name plus event_id (Meta, handling duplicate events). The TikTok Events API provides the same kind of server-side connection (TikTok, Events API).

Google Ads has two further reinforcements: enhanced conversions improve the accuracy of conversion measurement (Google Ads Help: enhanced conversions), and offline conversion import sends back sales by the GCLID of each click (Google Ads Help: offline conversion import). When a server-side sale has no matching front-end event, the GA4 Measurement Protocol can fill it in, linked to front-end interactions by client_id (Google, Measurement Protocol).

Consent and personal data should be settled before launch. Google's Consent Mode v2 adds two parameters, ad_user_data and ad_personalization, on top of ad_storage and analytics_storage (Google, Consent Mode). Article 20 of Taiwan's Personal Data Protection Act limits non-government agencies to using personal data within the scope necessary for the specific purpose of collection, and Article 19 lists consent as one of the bases for collection (Personal Data Protection Act). Article 5(3) of the EU ePrivacy Directive requires clear and comprehensive information to be provided before information on terminal equipment is accessed (EU ePrivacy Directive, 2002). These are pre-launch checks, not legal advice, and the scope that applies should be confirmed with legal counsel.

Merging social data and site conversions into one table

Video view data on each platform and conversion events on the site live in different systems. Comparing two versions takes one table aligned by creative ID and date, with one row per creative and date and columns for views, retention, add-to-cart, orders and revenue.

There are two merge keys: the creative ID (utm_content on the site, and the ID registered at upload on the platform) and the date. On the site side, use the GA4 BigQuery export, which is a full daily export and can also stream the current day's data (Google Analytics Help: BigQuery Export). On the platform side, pull through each API daily and write to the same dataset.

Metric names and definitions differ across platforms, so the numbers cannot be added together directly. Store platform metrics unchanged and build a separate mapping layer that records each field's definition and collection date, so that nobody later treats Instagram's average watch time and YouTube's retention rate as the same thing.

Decisions: allocate traffic by conversion, treat clicks as a reference

Traffic allocation is based on conversions and revenue, and clicks serve only as a reference, for three reasons. YouTube moved its ranking objective from click-through rate to expected watch time, because optimizing for clicks alone promotes clickbait videos that users do not finish (Covington et al., 2016). Optimizing an imperfect proxy metric past a critical point makes performance on the true objective worse (Karwowski et al., 2023). Recommender-systems research also lists optimizing for long-term outcomes as an open problem, since short-term engagement alone is not enough (Stray et al., 2022).

The main difference between allocation algorithms is whether traffic is allocated while exploring, or in fixed proportions until the test finishes. A fixed A/B test suits cases with few variants, enough samples and a need for a clear conclusion. A bandit shifts traffic toward better-performing variants during the test itself and suits cases with many creatives and a content pool that changes often. LinUCB uses a contextual bandit to explore while choosing, for news recommendation where the content pool keeps changing and traditional collaborative filtering does not apply (Li et al., 2010). In empirical results on advertising and news recommendation, Thompson sampling performed better than or as well as methods such as UCB, and the authors argue it should be a standard baseline (Chapelle and Li, 2011). A traditional A/B test can optimize only one aspect of a page at a time, and Amazon used a bandit to optimize combinations of several elements at once (Hill et al., 2017). For video, that corresponds to combinations of opening, product shot, caption card and call to action.

Platforms also have built-in allocation. In Meta's asset_feed_spec, multiple sets of images and text can be supplied and the system automatically tests and serves different variants, which is dynamic creative (Meta, Ad Creative reference). Performance Max has an asset report that lists each asset used in a campaign and allows performance to be compared (Google Ads Help: asset reporting). Conversion gains published by Meta, such as +13% for related video assets (Meta Advantage+ Creative) and a 20% lower average cost per result for similar merchants (Meta Advantage+ sales campaigns), are self-reported by the platform with no stated baseline, period or method, and serve only as a directional reference.

The order of judgment in each round is shown in the diagram below. First confirm the sample size has reached the minimum set in advance, then check guardrail metrics such as unsubscribes, negative reviews and returns, and only then compare conversions and revenue. If the sample is too small, keep the current allocation. If clicks are high but conversions are low, first check whether the video's promise matches the product page before adding traffic.

Production: how results return to the next batch of videos

For results to return to production, the video first has to be split into replaceable components: the opening, product shot, caption card and call to action are each stored separately. Each week, look at the merged data, rewrite or retire the weak components, feed the strong ones back into the asset library, and have the next batch redo only the parts that need changing.

On the copy side, a model can generate candidates, but a candidate is not a result. One study trained ad-copy generation on online click feedback while also noting that model-generated copy is not guaranteed to get a higher click-through rate than human-written copy (Chen et al., 2025). Another learned to generate and rank rewritten creative versions from historical A/B data (Mishra et al., 2020). The practice is to send every candidate through the same experiment and compare the candidates on actual feedback. Characters and scenes must stay consistent across versions, as described in Character and Scene Consistency in AI Video.

Automated publishing and data retrieval rely on each platform's API. The YouTube Data API's videos.insert can upload a video and set its metadata programmatically (Google, videos.insert), and the YouTube Analytics API provides real-time queries and custom reports (Google, YouTube Analytics API). The Meta Marketing API creates an ad creative object through the adcreatives endpoint and passes its id when the ad is created (Meta, Marketing API: Ad Creative), and results are returned by the Ads Insights API (Meta, Ads Insights API).

Attribution is not incrementality

The orders in the creative performance table are the result of attribution, which is not the same as incremental orders the ads brought in. GA4's data-driven attribution distributes the contribution of each interaction based on the account's own data (Google Analytics Help: attribution). What it assigns is credit, not causation, and people who would have bought anyway are counted too.

Answering how many fewer orders there would be without this video takes an experiment. Conversion Lift in Google Ads measures the incremental conversions that ads bring (Google Ads Help: Conversion Lift), and geo experiments compare regions as units, proposed by Vaver and Koehler (Vaver and Koehler, 2011). At the budget level, open-source marketing mix models can be used: Google's Meridian (Google, Meridian) and Meta's Robyn (Meta Marketing Science, Robyn). These methods need enough conversion volume and time, so run these tests monthly or quarterly to calibrate the daily automatic allocation.

The limits of automation: two human checkpoints

The first checkpoint is before launch, to review brand, facts and regulations. Automatically created creative still has to pass platform review: every Meta ad is reviewed against policy when the order is placed (Meta, Advertising Standards), and videos exported from TikTok Symphony Creative Studio get an AI-generated label under ad policy and must pass review again before publishing (TikTok Symphony Creative Studio Help). Platform review does not replace an in-house review of the video, and prices, efficacy claims and product facts must be checked in the internal process first.

MAP 3.5 of NIST AI RMF 1.0 calls for defining, assessing and documenting human oversight processes (NIST, AI RMF 1.0), which maps to a pre-publication review point, and the reviewer, the checklist and the result should all be recorded. Among the generative AI risks listed in NIST AI 600-1, information integrity refers to producing large amounts of content that does not distinguish fact from opinion (NIST, AI 600-1, 2024), and the more output is automated, the more this step needs people.

Disclosure requirements for AI-generated content differ by platform. YouTube requires disclosure of realistic-looking AI-generated or altered content, such as generating a realistic scene that did not happen, while fantasy or clearly unrealistic content, beauty filters and color adjustment, and script brainstorming do not need disclosure (YouTube Help: disclosing altered or synthetic content). Meta adds an "AI info" label to video and images, based on industry signals or self-disclosure by users (Meta Newsroom, 2024), and ads made or substantially edited with Meta's own generative AI creative tools show "AI info" under "About this ad" (Meta Newsroom, ad transparency). TikTok requires labeling of realistic AI-generated content and uses C2PA Content Credentials to label AIGC from other platforms (TikTok Newsroom, 2024), and after its November 2025 update it adds an invisible watermark to uploads that carry C2PA credentials (TikTok Newsroom, 2025).

The second checkpoint is a weekly review of the results of automatic adjustments, pausing and rolling back anything that deviates from expectations. The system retrieves data and allocates traffic every day, while the team reviews videos weekly and verifies incrementality monthly, as divided in the diagram below.

Rollout order

• Connect tracking and consent first. What to do: put creative IDs and utm_content live, have purchase carry transaction_id and items, and finish the Consent Mode parameters and consent banner. Done when: ten randomly picked orders can each be traced back to a creative ID, and Pixel and server-side events match in count after deduplication by event_id.
• Build the creative performance table and have people read it weekly. What to do: write platform metrics and the GA4 export to one dataset daily and merge by creative ID and date. Done when: a complete creative-by-date table has been produced for four weeks running, and the team can say from the table which components to keep or retire in the following week.
• Allocate by rule. What to do: set the minimum sample size, the traffic adjustment range and the guardrail metrics first, then allocate with a bandit or platform dynamic creative. Done when: the rules are written down, traffic can be paused and rolled back when a guardrail triggers, and the rollback has been rehearsed at least once.
• Let production generate candidates automatically, and publish after human review. What to do: replace only weak components and send candidates through the same experiment. Done when: every published version has a review record covering brand, facts, regulations and AI disclosure checks, and incrementality experiments produce results each month or quarter that can be used to recalibrate.

When this does not fit

When conversion volume is too low, statistical results are unstable and a bandit's allocation can be thrown off by noise, so start with tracking and manual reading, and move to rules once volume is enough. For product categories where product information or prices change often, or that are regulation-sensitive, automated output should keep more human review, and the review frequency should not be decided by performance data alone.

Without a company-owned website, or when checkout happens on a third-party platform and order data is unavailable, conversions cannot return to the creative performance table and only in-platform metrics are available, so most of the attribution and incrementality methods in this article do not apply.

POLISH can plan everything from video production through site tracking to data integration. For a worked example of video production, see how the animal welfare awareness short film project was done. For the service itself, see AI video production, and to start a project, contact POLISH.