Every performance marketer running paid campaigns on Xiaohongshu eventually faces the same wall: a learning phase that drags on, inflated cost-per-lead numbers, and an algorithm that seems to be burning budget without delivering results. It is one of the most frustrating early-stage experiences on the platform, and it catches even experienced media buyers off guard because Xiaohongshu’s ad system behaves quite differently from Meta or Google.

The platform now hosts over 300 million monthly active users, the majority of them higher-income Chinese consumers actively searching for product recommendations, lifestyle content, and brand discovery. For brands targeting this audience, Xiaohongshu advertising is not optional — it is essential. But the path from campaign launch to efficient lead generation runs directly through the algorithm’s learning window, and knowing how to navigate it can be the difference between a campaign that scales and one that gets abandoned after the first month’s spend.

This article breaks down exactly what the Xiaohongshu ad learning phase is, why it matters for cost-per-lead efficiency, and a practical seven-step sprint framework you can implement immediately to exit the learning phase faster and drive sustainable CPL improvements.

What Is the Xiaohongshu Ad Learning Phase?

When you launch a new ad set on Xiaohongshu’s advertising platform (known as Juguang), the system enters a calibration period during which it tests delivery across different audience segments, placements, and creative combinations to determine where your ad performs best. This is the learning phase. During this window, the platform’s machine learning model is actively gathering data on who clicks, who converts, and what signals correlate with your target action — whether that is a form fill, a message inquiry, or a click-through to your landing page.

Unlike Meta Ads, which requires roughly 50 optimisation events per ad set per week to exit learning, Xiaohongshu’s threshold and mechanics are less publicly documented, making it harder for advertisers unfamiliar with the ecosystem to know whether they are progressing or stagnating. Generally speaking, you should expect the algorithm to need somewhere between 100 and 200 optimisation events before delivery stabilises. During this period, CPL figures will be erratic, and it is normal to see wide swings in performance from day to day.

Why the Learning Phase Kills Your CPL Before the Campaign Begins

The core problem is that most advertisers judge campaign performance too early. They see high CPLs in the first week, panic, make structural changes to the campaign, and inadvertently reset the learning phase entirely. Every time you significantly alter a campaign’s budget, audience targeting, creative, or bidding strategy, the algorithm treats it as a new signal environment and restarts its calibration process. What was a campaign two days from stabilisation suddenly becomes a campaign that is starting from scratch.

This creates a cycle where marketers are perpetually stuck in the most expensive period of any campaign — the exploratory phase — without ever reaching the efficient delivery window where the algorithm has learned enough to serve ads to the highest-value users at the lowest possible cost. For brands operating on finite monthly budgets, this pattern can consume an entire quarter’s ad spend without a single meaningful data point about sustainable CPL.

Five Signals You’re Stuck in the Learning Phase Too Long

Before running a CPL optimisation sprint, it helps to confirm you are actually dealing with an extended learning phase rather than a targeting or creative problem. Watch for these indicators:

• Volatile daily CPL with no downward trend after 10 or more days of continuous delivery
• Low impression volume relative to your target audience size, suggesting the algorithm has not found a reliable delivery cohort
• Flat or declining click-through rate across all creatives without clear differentiation between ad variants
• Conversion tracking showing zero or near-zero events despite visible clicks to your destination page
• Budget under-delivery where the platform is not spending your daily budget ceiling, indicating low confidence in the algorithm’s targeting signals

If you are seeing three or more of these signals beyond the two-week mark, you are almost certainly in an extended learning phase that requires deliberate intervention rather than passive optimisation.

The CPL Optimisation Sprint: A Seven-Step Exit Framework

A CPL optimisation sprint is a structured, time-boxed process designed to exit the learning phase efficiently and establish a CPL baseline that supports scaling. Here is how to execute it:

1. Consolidate your campaign structure – Fragmented campaigns with too many small-budget ad sets starve the algorithm of data. Consolidate into fewer, better-funded ad sets to accelerate event volume accumulation.
2. Set a realistic learning budget – Calculate your estimated CPL target and multiply by at least 50 to determine your minimum learning budget. If your target CPL is 200 RMB, you need at least 10,000 RMB available for the learning phase before optimising.
3. Audit conversion event setup immediately – Many campaigns extend the learning phase simply because conversion tracking is broken or misconfigured. Verify your Juguang pixel is firing on the correct thank-you page or event trigger before spending another renminbi.
4. Freeze structural changes for 14 days – Commit to a change moratorium. Minor creative refreshes within the same ad set are acceptable, but do not alter budgets, audience targeting, or bid strategies during the learning window.
5. Launch three to five creative variants from day one – Give the algorithm options to explore. Creative diversity helps the system identify resonant content faster than a single-creative ad set, reducing time to stable delivery.
6. Use a conversion-optimised bid strategy over manual bidding – Manual CPC bidding forces the algorithm into a cost-first mindset that can conflict with conversion optimisation. Switch to oCPX (optimised cost per action) bidding to align the system’s incentives with your actual goal.
7. Set a bid cap that is 20 to 30 percent above your target CPL – Bidding too close to your target CPL causes the system to restrict delivery to only the safest predicted conversions, limiting exploration. A slightly elevated cap gives the algorithm room to learn while still anchoring toward your efficiency goal.

Creative Velocity: Why UGC-Style Content Shortens Learning Time

Xiaohongshu’s native content environment is dominated by user-generated content — personal reviews, lifestyle photography, unboxing videos, and authentic product demonstrations. Ads that visually and tonally blend with organic content consistently outperform polished brand productions in both engagement and conversion metrics on the platform. More importantly for the learning phase, higher engagement rates give the algorithm stronger early signals to work with, which accelerates the calibration process.

Investing in UGC-style creatives before launch is one of the highest-leverage moves you can make. This means shooting content on mobile devices, featuring real users or relatable talent rather than models, writing captions that mirror how genuine Xiaohongshu users describe products, and leaning into platform-native formats like the vertical image grid and short-form video notes. Brands that integrate their influencer marketing output directly into their paid creative pipeline often see the fastest learning phase exits, because the content has already been validated by organic engagement before it goes into media buying rotation.

Audience Structure Mistakes That Extend the Learning Phase

Audience targeting errors are the second most common cause of an extended learning phase, after conversion tracking failures. The most frequent mistake is over-segmentation: building hyper-specific interest stacks that restrict the available audience to a size too small for the algorithm to find meaningful conversion patterns. On Xiaohongshu, broader audience definitions during the learning phase are almost always preferable to narrow ones, because the algorithm’s own machine learning is the most effective segmentation tool you have access to.

Another common error is layering too many exclusion conditions on cold audiences from the start. While exclusion targeting makes sense at scale, applying it prematurely reduces the event pool the algorithm needs to learn from. Start with interest-based or keyword-based targeting using moderate parameters, let the algorithm identify your highest-converting audience segment, and then introduce exclusions and lookalike expansion once you have exited the learning phase and established a stable CPL. If your brand is working with a specialist Xiaohongshu marketing partner, they should be able to provide category-level benchmark data on audience sizing norms for your vertical.

Budget Floors and Bid Strategies That Help the Algorithm Learn Faster

One of the least discussed levers in Xiaohongshu ad optimisation is the relationship between daily budget floors and learning speed. The platform’s delivery algorithm distributes budget across a rolling optimisation window, and campaigns with budgets set too close to the minimum daily threshold often see uneven delivery patterns that distort learning signals. As a practical guideline, set your daily budget at a minimum of three to five times your target CPL to give the system enough room to explore delivery patterns without budget-pacing constraints interrupting the flow.

On bid strategy, the oCPX model is specifically designed to reduce CPL over time by using historical conversion data to predict which auctions are most likely to generate a qualifying action at or below your target cost. During the learning phase, this model is actively building its prediction framework, which is why maintaining a consistent bid cap (rather than adjusting it frequently) is critical. Every bid cap change is interpreted by the system as a new cost constraint, potentially triggering re-exploration of delivery segments the algorithm had already begun to optimise. Patience during this window is a strategic discipline, not passive management.

Measurement Setup: Tracking Conversions the Right Way on Xiaohongshu

Conversion tracking on Xiaohongshu runs through the Juguang pixel, which can be implemented via direct code integration on your landing page or through supported tag manager configurations. The most common tracking failure is firing the conversion event on the wrong page action — for example, tracking a page view on the lead form page rather than a submission event on the confirmation page. This floods the algorithm with false positive signals and teaches it to optimise for users who view the form, not users who complete it, which produces the worst possible combination: high event volume, low actual lead quality, and an artificially low reported CPL that does not reflect real business outcomes.

Before launching any campaign sprint, conduct a full pixel audit using Juguang’s built-in event testing tool. Verify that each tracked event fires exactly once per qualifying action, that the event category matches your optimisation goal, and that the conversion window aligns with your typical sales cycle. For brands managing complex multi-touch conversion paths, integrating your Xiaohongshu campaign data with a broader analytics framework (CRM attribution, for instance) ensures that your CPL optimisation decisions are grounded in revenue-correlated data rather than platform-reported metrics alone. Hashmeta’s AI marketing services include conversion infrastructure setup as part of campaign onboarding, which significantly reduces time-to-learning for new advertisers on the platform.

When to Bring in a Specialist Xiaohongshu Marketing Partner

The learning phase is manageable for teams with hands-on Xiaohongshu advertising experience, but for brands entering the platform for the first time or scaling aggressively, the cost of trial-and-error learning can be prohibitive. A specialist partner brings two immediate advantages: historical benchmark data across comparable campaigns, and the structural confidence to commit to a learning window without mid-flight panic adjustments that reset progress.

Beyond campaign mechanics, a strong Xiaohongshu marketing agency should be able to integrate your paid media strategy with your organic content programme, influencer seeding efforts, and broader content marketing calendar. Xiaohongshu’s algorithm rewards brands that maintain genuine platform presence, and organic engagement signals can meaningfully improve paid ad quality scores over time. If your goal is sustainable CPL reduction at scale, the paid and organic strategies need to be developed in parallel rather than in isolation.

For brands looking at broader performance marketing infrastructure across Southeast Asia and China, Hashmeta’s integrated approach, which spans AI marketing, influencer discovery through StarScout AI, and full-funnel campaign management, means you are not just optimising a single platform in isolation. You are building a regional digital presence that compounds in value over time.

The Learning Phase Is a Feature, Not a Bug — If You Work With It

The Xiaohongshu ad learning phase is not an obstacle to be avoided. It is the algorithm’s investment period, during which it builds the predictive model that will eventually deliver your leads at the lowest sustainable cost. The mistake most advertisers make is treating it as a performance problem rather than a structural process, which leads to the reactive decisions that keep campaigns perpetually stuck in the most expensive delivery window.

By consolidating campaign structure, protecting your conversion tracking, committing to creative diversity, and maintaining disciplined bid management through the full learning window, you give the algorithm everything it needs to exit calibration quickly and start driving efficient results. The brands that master this process on Xiaohongshu are the ones that build durable CPL efficiency advantages that are very difficult for competitors to replicate without going through the same learning investment.

Start with the seven-step sprint framework outlined here, audit your tracking setup before anything else, and give the algorithm the data volume it needs to work. The exit from the learning phase is closer than most advertisers realise — it usually just requires getting out of the algorithm’s way.