tokconnectBrand-keyword research shows how people talk about products in public posts. It answers a different question from an audit of a competitor’s own account. “What is Starbucks doing?” is too broad for search results. “What order language appears in selected relevance-ranked third-party posts returned for Starbucks coffee and Dunkin coffee?” states the sample and the level of claim. The completed output should show source links, fields, observations, alternatives, and a test the brand can support.
Fix the comparison before collecting records
Choose two competitors or category references, matched seed phrases, a sort order, result limit, date, and public fields to save. The saved coffee example uses separate relevance searches for Starbucks coffee and Dunkin coffee, with selected third-party posts. It does not claim that either post came from a brand-owned account or represents each brand’s strategy.
Use the same query shape where possible. If you must use two brand names, preserve both terms. Then run a generic query such as coffee order to see whether a phrase may be category language rather than brand-specific language. TikTok’s official Research API documentation also distinguishes public video and account queries and explains returned paging values for approved clients; it is a separate approval-based program from TokConnect’s research workflow.
Compare the saved records
The saved September 18 evidence selected a Starbucks-related post by @julianaa_braga. Its stored caption says “this is a must try order…” and lists 9,124,160 views, 931,136 likes, 2,607 comments, and 62,111 shares. The Dunkin query selected @jesi02_, whose stored caption says “This Dunkin coffee order is a must-try!! …” and lists 913,435 views, 52,551 likes, 437 comments, and 4,927 shares.
| Query | Selected public post | Caption observation | Stored public counts |
|---|---|---|---|
| Starbucks coffee | @julianaa_braga | “must try” order language | 9,124,160 views; 2,607 comments |
| Dunkin coffee | @jesi02_ | “must-try” order language | 913,435 views; 437 comments |
The counts describe selected records from different searches. They do not measure a fair performance contest: topics, account audiences, post age, and ranking conditions differ.
Test the category alternative
A relevance-ordered coffee order search included the Starbucks-related record and a general post from @megahfun_, with stored caption “My coffee order,” 1,964,547 views, 198,537 likes, and 497 comments. That makes one limited interpretation possible: recommendation language around an order appears in more than one coffee-related result. The sample does not prove that Starbucks or Dunkin created, owned, or caused the pattern.
Before you name a format, watch the linked public posts. Metadata alone cannot establish visual sequence, soundtrack, edits, spoken claims, or product handling. Write reviewer observations in their own column and preserve the reviewer and date.
Turn an observation into a supportable test
A café can test an order explanation only if staff can verify the drink, substitutions, price language, and availability. The proposed angle is “explain one order and the trade-off it solves,” not “copy a must-try script.” A coffee equipment brand might reject it because an order recommendation does not demonstrate its product. Then use a making-coffee query and repeat the sample protocol.
Compare [brand A] and [brand B] using [sort] searches on [date]. Save the exact query, selected public video URL, caption, and counts. Run one generic category query to test whether the observed language belongs to the category. Do not infer account ownership or strategy from third-party posts. Return one supportable content test and the proof it needs.
Normalize the selected public counts
Ratios make the two stored records easier to inspect. For the Starbucks-related post, likes per view are 931,136 ÷ 9,124,160 = 10.21%; comments per view are 2,607 ÷ 9,124,160 = 0.029%; shares per view are 62,111 ÷ 9,124,160 = 0.681%. For the Dunkin-related post, the equivalent figures are 52,551 ÷ 913,435 = 5.75%, 437 ÷ 913,435 = 0.048%, and 4,927 ÷ 913,435 = 0.539%.
These calculations describe public totals on two selected posts. They do not explain the differences or identify a winning creative choice. They do identify a useful follow-up: the Dunkin-related record has more comments per view in this pair, so a researcher can collect the same number of comments from each post and compare question types without treating comment count as customer demand.
| Signal | Observed sample result | Hypothesis | Test design |
|---|---|---|---|
| Caption language | Both brand-query posts use “must try” order language | Specific orders may be a category hook | Code ten relevance-ranked public posts per query for order framing |
| Generic alternative | Generic coffee-order query includes order framing | Language may be category-wide | Compare generic and brand-query shares using the same coding rubric |
| Comments per view | 0.048% vs 0.029% in this pair | Questions may differ by post | Retrieve 20 comments per selected post; code question, request, objection, and praise |
Use an opportunity matrix after coding: keep an angle when it repeats in the matched sample, fits the client’s product, and the client can verify its claims. Reject it when it only appears in an isolated result, requires unavailable product facts, or depends on copying a named creator. A café can test an order explanation with a verified drink and substitutions; a coffee-tool brand should use a preparation query instead.
Hand off the brief
Question: Does recommendation language appear in selected coffee-order posts?
Sample: one selected relevance result per brand query; one generic coffee-order query
Observed: both selected brand-related captions use “must try” language
Alternative: generic coffee-order result also uses order framing
Limit: third-party posts; no controlled performance comparison
Decision: validate one source-backed order explanation for a café
Next validation: reviewer watches links; location verifies the order detailsExpand the sample only after the client accepts the question and scope. A larger unstructured pile of links produces less useful analysis than a smaller repeatable record.