When a Summary Replaces an Answer—and When It Sells a Conversation
The Reddit evidence shows that eligibility for AI-generated visibility can increase participation, but an average effect does not explain why. This chapter focuses on the value a user is trying to obtain from a community. When a concise summary delivers most of that value, it can replace a visit. When the value lies in perspectives, disagreement, advice, or interaction, the same summary may make the source more worth visiting. That distinction gives us a way to interpret the different responses inside Reddit without treating either factual or experiential content as an automatic winner.
Two kinds of value in one community ecosystem
The study adapts a distinction from information economics: search goods and experience goods. A search good is something whose quality can be assessed before consumption or from relatively direct information about it. Applied to online content, the category includes factual, verifiable material—a question whose answer can be checked, a reference point, or information that loses little when compressed into a short explanation.
An experience good is different. Its value becomes clear through consumption. In an online community, that can mean reading how people describe a situation, comparing several opinions, seeing advice challenged, or following the interaction that develops in a comment thread. The value is not only the proposition stated in a post. It is also the texture of the discussion, the credibility that comes from lived experience, and the ability to ask a follow-up question of people rather than of a summary generator.
The distinction is not the same as “easy content” versus “hard content.” A technical question can require a long discussion, while a personal question can sometimes receive a concise answer. The more useful test is what the user would lose if the content were reduced to a summary. If most of the value survives compression, the content behaves more like a search good. If the value depends on seeing the full exchange, it behaves more like an experience good.
This way of classifying content matters because AI search changes the source’s competitive position after the summary appears. The summary is not simply another link. It is a partial consumption of the source. Whether that partial consumption satisfies the user or creates curiosity depends on the source’s residual value—the part that remains unavailable, unconvincing, or incomplete on the results page.
Consider the difference between a request for a sports score and a request for advice about a difficult relationship. A concise, accurate score may complete the first task. A concise description of common relationship advice may help with the second, but it cannot show how different people interpret the situation, what assumptions they challenge, or how the person asking the question responds to follow-up. The examples are not meant to divide all online material cleanly. They show why the same amount of summarization can have different competitive effects depending on what the user is trying to consume.
The distinction also has a supply-side implication for community platforms. A forum does not create value only by storing posts that can be retrieved. It creates value when people contribute experience, respond to one another, and make a topic legible to someone who did not know what to ask at the outset. AI search can expose that accumulated value, but it may also extract the most portable factual layer and leave the platform to carry the cost of hosting the interaction. The balance depends on whether new discovery brings enough participation to sustain the part of the ecosystem that summaries cannot reproduce.
That is why a platform should not ask only whether its content is “visible to AI.” It should ask which layer of its value the interface is making visible, which layer it is consuming on the results page, and which layer still requires a community visit. Those questions are more informative than a binary label about whether AI search is a partner or a competitor.
The residual need after a static summary
A static AI summary can push the user in two opposite directions. For search-like content, it may answer the question directly. The user may have no strong reason to open the cited page, because the summary has already delivered the information they were seeking. That is a replacement path.
For experience-like content, the summary may perform a different function. It can tell the user that a relevant community exists and point toward a conversation that cannot be fully represented in a few sentences. The summary becomes a discovery path. It reduces the cost of finding the source without eliminating the reason to consume it.
The two paths are not properties of the model alone. They result from the interaction between the interface and the source. A particularly clear summary can strengthen replacement even when the underlying material is rich. A thin or ambiguous summary can leave a factual question unresolved. Likewise, a community that is nominally experience-oriented may offer a discussion so repetitive or uninformative that a summary is enough for a particular user. The categories describe a tendency in how value is consumed, not a permanent label attached to every query.
The distinction also changes how “discovery” should be understood. Discovery is not simply the first impression or the first click. It is the process by which a source enters a user’s consideration set and becomes a plausible place to spend attention. AI search can widen that set by surfacing a niche community in response to a query. Whether the community benefits then depends on what the user finds after arriving: a useful exchange, a credible answer, a welcoming interaction, or some other form of value that the summary did not supply.
This is why the Google–Reddit result cannot be interpreted from the number of citations alone. A citation can be a substitute for a visit, an advertisement for a visit, or both for different users and queries. The content-value distinction supplies the missing mechanism: it asks what the summary is capable of doing and what the source still needs to do for the user.
What the content comparison shows
To test this mechanism, the researchers classify Reddit communities as search goods or experience goods and compare how the AI Overview exposure effect differs between them. The classification uses each subreddit’s name, title, and public description. Gemini 3 Flash applies a rule based on whether a user could obtain most of the community’s value by reading posts alone, without comments or interaction. The model returns a label, a confidence score, and a short justification. Of the 105,012 communities in the main cohort, 104,975 receive usable classifications: 59,450 are labeled search goods and 45,525 experience goods.
The analysis then asks a more demanding question than whether the two categories have different average activity. It uses a triple-difference comparison to estimate whether the gap between AI-eligible and policy-excluded communities is larger for experience goods than for search goods. This matters because experience-oriented and factual communities may differ even without AI search. The relevant signal is the additional SFW-versus-NSFW change associated with content type.
The logic has three moving parts. First, activity is compared before and after the AI Overview exposure change. Second, the change for SFW communities is compared with the change for NSFW communities. Third, that exposure contrast is compared across the two content classifications. The third comparison is what asks whether experience-oriented communities receive an extra lift beyond the general difference associated with being an AI-eligible Reddit community. Without that third step, a larger experience-community average could be mistaken for evidence that the summary caused the difference.
The result supports the proposed mechanism. The experience-good treatment effect is 2.3 times the search-good effect for comments and 2.8 times the search-good effect for comment authors. In other words, the exposure contrast is positive in both categories within Reddit, but the relative gain is substantially larger where the value of the community depends more on discussion and lived experience. The result fits the idea that a static summary can surface experience-based material without fully replacing it.
The ratios should be read as relative comparisons, not as a universal multiplier for every platform. They describe how the estimated treatment effects differ inside the study’s Reddit sample and under its classification rule. They do not mean that an experience-oriented community will always receive 2.8 times as many new participants, or that a factual platform will lose a fixed share of visits. The value of the ratio is explanatory: it shows where the discovery channel is strongest in this setting and supports the claim that interaction-dependent value is harder for a static summary to absorb.
The two categories’ positive direction is important. A simplified retelling might say that AI search helps experience goods and harms search goods. That is not what this Reddit study finds. Both kinds of communities show positive effects in the platform’s forum setting; the experience-oriented communities show the stronger response. Reddit’s community structure may allow even factual discussions to benefit from the act of finding a relevant group, asking a follow-up question, or seeing how other users apply information. The search-versus-experience distinction predicts the relative size of the effect better than it predicts a universal sign.
The margin also matters. A larger experience-good effect for comment authors means more distinct people enter the discussion, not merely that a small circle of regulars produces more words. A larger effect for comments means more discussion volume. The study reports both margins so that “engagement” does not hide whether the change is participation, intensity, or both. The content mechanism is therefore tied to observable behavior rather than left as a metaphor about richness.
A comparison, not a taxonomy of winners
The classification is informative but not a direct measurement of every community’s actual content. It is based on metadata—names, titles, and descriptions—rather than a complete reading of every post and comment. The labels come from a model-assisted rule, and a small number of communities are not classified because the response is blocked or cannot be parsed. The researchers report that the heterogeneous result remains when attention is restricted to high-confidence classifications, but confidence does not turn the labels into randomized treatment assignments.
That boundary affects the interpretation in two ways. First, a model may infer the purpose of a community from a description that is incomplete, outdated, ironic, or written for insiders. Second, the classification is fixed at the community level while users encounter individual posts and threads. The analysis is therefore strongest as evidence that communities with more experience-oriented metadata responded differently on average, not as proof that every post followed the same replacement or discovery path.
There is another reason not to read the categories as fixed identities. Communities can contain both answer-like and experience-like material. A subreddit may host factual explanations in one thread and deeply personal advice in another. The same post can also serve different needs: one reader wants a quick fact, while another wants the reasoning and disagreement around it. The relevant unit is often the query–content encounter, not the community name alone.
The study’s broader implication is therefore conditional. Factual-content platforms face more substitution pressure when an AI system can satisfy the information need directly. Discussion-oriented platforms may gain when the summary functions as a signpost to richer material. But neither statement is a universal verdict about a platform category. Content composition, user intent, moderation, interface design, and the outcome being measured all change the balance.
This conditional view also makes the finding useful for the next change in the search interface. The replacement-versus-discovery mechanism depends partly on what a static summary cannot do. If the search page becomes conversational and can handle follow-up questions, it may absorb more of the depth that previously required a community visit. The content distinction tells us what to look for; it does not tell us that the boundary will stay fixed.
The most defensible conclusion is that AI search does not confront community platforms with one effect. It changes the value of being found according to what users still need after seeing the answer. In Reddit, both factual and experience-oriented communities gained participation under the observed exposure contrast, but the experience premium was much larger. To use that result correctly, we now need to combine three elements—what the interface delivers, what the source uniquely supplies, and which outcome the evidence measures—before asking how a more interactive interface changes the relationship.