A new Instagram research paper raises a practical question for marketing leaders: how closely does an AI system’s test resemble the decision you are about to make?
A shortlist arrives with three creators, three prices, and three confident scores. One looks like the obvious choice. The number is high, the explanation is polished, and the deadline is close.
Before approving the spend, ask a less comfortable question: what did that score have to get right before it earned a place in this decision?
Perhaps it summarized historical engagement. Perhaps it predicted reactions to content that already existed. Perhaps it ranked creators against a reference group. Each can be useful. None, by itself, establishes what a new collaboration will contribute to this brand, at this price, under this brief.
The commercial value of AI depends partly on how much work it removes. It also depends on whether the evidence behind its recommendation survives the move from a demonstration to an actual buying decision.
What the new research actually tested
Research in focus · October 2026
An October 5 arXiv preprint by Pedro Victor de Sousa Lima and Olga Goussevskaia examines Instagram engagement across account types. Its working corpus contains roughly 15.7 million posts from 766,000 classified accounts, with data ending in 2024. Modeling uses samples of up to 50,000 posts per segment.
The models predict likes, comments, and engagement rate from post characteristics, excluding follower count and performance history as predictive inputs. On log-transformed likes, the reported R² reaches 0.74 for event and creative-work accounts, versus 0.33 for people.
The test randomly divides posts 80/20, allowing the same account into both sets. The authors acknowledge limits to generalization, incomplete sponsorship detection, imperfect account classification, and a correlational design. This is engagement research, not a sales experiment or a validation of any commercial platform.
R² describes fit against a mean-prediction baseline for the measured target. It is not a percentage probability that a campaign will succeed. The manuscript is treated here as a preprint; independent replication was not verified.
The buying lesson is our interpretation: a test result becomes commercially useful when the buyer can explain which part of the proposed decision it supports. A large dataset and an impressive metric can begin that explanation. They cannot finish it.
Define the decision before evaluating the score
Consider three requests a marketing team might make of the same tool. Find promising creators in an unfamiliar category. Estimate the likely response to a proposed sponsored video. Decide whether an existing partnership deserves a renewal.
Those requests need different evidence. Discovery needs a useful way to find and compare unfamiliar candidates. A content forecast needs information about the planned execution. Renewal needs the actual partnership history, including what the brand paid and what it received.
A single number can help organize those questions, but it should not quietly replace them. If a tool was evaluated for finding relevant candidates, treat its recommendation as a shortlist input until additional evidence supports the budget decision.
Write one sentence before the demo: “We need help deciding whether to commission this creator for this deliverable, in this market, within this budget.” That sentence makes it harder for an attractive but irrelevant metric to win the meeting.
Ask for a test that resembles the work
NIST’s AI Risk Management Framework 1.0 connects accuracy reporting with realistic test conditions, documented methodology, and results for relevant data segments. It also calls for continued assessment after deployment. This is general guidance, not an influencer-marketing certification.
Applied to creator selection, that principle suggests a practical set of questions. They are proposed purchasing checks, not results from the Instagram study.
| If you need to… | Ask to see… |
|---|---|
| Discover unfamiliar creators | Evaluation on creators excluded from model development. |
| Plan the next campaign | Predictions recorded before later outcomes became available. |
| Buy in a particular market | Results and coverage for the relevant language, category, and format. |
| Commit a larger budget | The errors around the approval threshold, including costly overestimates. |
There is no requirement that every tool solve every row. A discovery assistant may deliver real value without forecasting revenue. The scope of the buying decision should match the scope of the evidence.
A vendor can also be candid about a gap. “We have not tested that market yet” is actionable information. It tells the buyer to narrow the initial use, collect relevant examples, or run a limited pilot. An unqualified score hides that choice.
Keep prediction and financial impact on separate lines
Google’s Meridian documentation makes a related distinction in marketing mix modeling: models can fit or predict the data well while producing different ROI estimates. Causal interpretation requires additional assumptions. That guidance concerns media measurement, not creator scoring, but the distinction matters here too.
Imagine a creator correctly predicted to attract substantial engagement. The product might still be unavailable where viewers live. The price could be uncompetitive. The brief might entertain without explaining the offer. Existing customers might account for many tracked purchases.
Those are illustrative possibilities, not findings about a specific campaign. They show why the business case needs more than an accurate audience-response estimate.
Ask the provider to name the outcome behind each score. Relevance, engagement, estimated views, attributed sales, and incremental profit should remain distinguishable. If the system offers a composite, ask which components are measured, which are inferred, and how missing inputs affect the recommendation.
FLONCI has previously examined the case for incrementality. The additional question here comes earlier: does the evidence used to validate the recommendation match the outcome being promised?
The expensive errors deserve their own review
An average error can obscure the mistakes that matter most to a buyer. Slightly underestimating an inexpensive experiment and substantially overestimating a major commitment do not create the same budget exposure.
A useful evaluation should therefore inspect decisions near the point where the team would approve more spending. Which recommendations looked strong enough to fund but later disappointed? What evidence was missing? Was the error a forecast problem, a poor match, a price problem, or an execution failure?
These questions should not become a demand for perfect prediction. They should help set sensible limits. A recommendation supported mainly by broad benchmarks might justify further investigation. A larger commitment may require directly relevant history and stronger commercial evidence.
The opposite error matters as well. An overly cautious system can repeatedly reject valuable unfamiliar creators. Record both missed opportunities and weak approvals. Otherwise, a tool can appear prudent simply by recommending very little.
Make uncertainty change the next action
A disclaimer at the bottom of a confident recommendation does little practical work. Useful uncertainty should affect the decision.
If the audience’s relevant geography is unknown, request verification before treating reach as commercially useful. If sponsored history is thin, consider a bounded test before committing to a longer relationship. If the proposed creative differs substantially from the material reviewed, assess the brief rather than assuming the historical pattern transfers.
Missing information should remain visibly missing. An uncollected metric is not a measured zero; an unavailable price is not evidence that a creator is inexpensive. The interface can stay simple while preserving these distinctions.
For example, a recommendation could explain the useful evidence, the important gap, and the next proportionate action in three short sentences. A buyer should be able to understand why the system recommends proceeding, testing, declining, or gathering more information.
This is a proposed standard for decision support. It does not imply that every platform already follows it, or that a particular label guarantees a good commercial result.
A hypothetical launch meeting
Imagine a retailer preparing a new skincare launch. The team has identified two creators with relevant content and broadly similar quotes. An AI tool ranks one higher, but its evidence is drawn mostly from entertainment-led organic posts. The second creator has fewer spectacular results and several clearly documented product demonstrations.
The team should resist making either creator the automatic winner. Organic success can contain valuable creative evidence. Sponsored examples can be informative without proving that the next execution will work. The task is to discover what each option can credibly contribute to this launch.
The buyer asks for the comparison to be restated around the actual brief: show the product’s use, explain a specific consumer concern, and reach people who can buy it. The team checks relevant audience information, the proposed claims, the deliverables, and usage rights.
If the evidence still cannot distinguish the options, the right result may be a limited test or a request for more information. A more detailed decimal score would not resolve an evidence gap.
The test’s purpose and evaluation are agreed before launch. It might examine whether the demonstration is understood and whether the intended audience responds. Any financial conclusion must respect the measurement design. No outcome is assumed in this example.
A small evidence pack is more useful than a bigger promise
Marketing teams should not need a machine-learning department to evaluate a recommendation. A concise evidence pack can answer the essential questions: intended use, data dates, relevant coverage, test method, typical errors, important limitations, and what happens when evidence is insufficient.
Ask for a comparison with a simple existing process, too. If the tool saves review time, assess that benefit directly. If it improves selection, agree how that improvement will be evaluated. If it claims financial impact, insist on an appropriate basis for that claim.
For an early pilot, predictions can be recorded before outcomes arrive while experienced people retain approval. Preserve the initial inputs and recommendations, then review what happened without rewriting the original expectation. A small pilot can reveal operational problems; it should not be presented as conclusive proof of broad ROI.
Keep the test proportionate to the decision. It is reasonable to learn gradually from a limited use case. It is much harder to justify scaling an unexamined recommendation simply because the output looks authoritative.
Where FLONCI fits
FLONCI’s decision-before-spend perspective makes this a standard to apply to ourselves as well as to the wider market. A recommendation should make the creator, the objective, the price, and the evidence easier to evaluate together.
This article does not claim that FLONCI has passed the tests described here or demonstrated a particular accuracy or savings rate. It sets out the evidence a brand should expect when AI is used to support an influencer investment decision.
That expectation can remain practical. Ask what the score measures. Check whether its test resembles the work. Make the important gaps visible. Let the strength of the evidence determine the size of the commitment.
Know before you spend includes knowing what the recommendation has actually proved.
Research notes and sources
Sources checked October 17, 2026. The purchasing framework and hypothetical scenario are FLONCI editorial analysis. No new experiment or commercial-system audit was conducted.
- Lima and Goussevskaia, Predicting Engagement with Sponsored Content Across Account Types on Instagram. October 5, 2026, arXiv v1.
- NIST, AI Risk Management Framework 1.0: AI Risks and Trustworthiness. 2023 framework. The live page notes that a revision is in progress.
- Google Meridian, About MMM as a causal inference methodology. Page updated September 1, 2026. Google is a commercial advertising-platform provider.
Related reading: Small Creators Need a Portfolio Strategy, Not an Engagement-Rate Shortcut.
No comments:
Post a Comment