Walmart's Polaris Algorithm: What Actually Ranks Products In 2026

The 2012 Announcement Everyone Still Quotes

On August 30, 2012, Walmart issued a press release announcing a new search engine for Walmart.com. It was named Polaris. It was built in about ten months by a small team inside the company's technology lab. It used semantic search to infer what a shopper meant rather than matching literal strings, so a query for denim could return jeans. Walmart reported an approximate 10 to 15 percent increase in shoppers completing a purchase after searching with it.

That release is still online, and it is the source almost every current article traces back to, usually without saying so. The claims in seller-facing content about Polaris understanding intent and mining synonyms are lifted, sometimes word for word, from a document written when the iPhone 5 had not shipped yet.

None of this means the concept was wrong. Semantic search really is how Walmart works, and the 2012 announcement was directionally correct about where retail search was heading. The problem is that treating a 2012 press release as a specification produces advice that is vague where it should be specific and specific where it should be current. You end up being told to use relevant keywords, which is true of every search engine ever built and tells you nothing about what to do on Tuesday.

What Actually Retrieves Your Product

Walmart's research team has published the architecture repeatedly. The paper Semantic Retrieval at Walmart describes a hybrid system deployed in production that combines a traditional inverted index with embedding-based neural retrieval, built specifically because tail queries have complex and specific intent that keyword matching handles badly. A follow-up paper, Enhancing Relevance of Embedding-based Retrieval at Walmart, states that the models were deployed in a hybrid retrieval setting serving all traffic on walmart.com and describes techniques including typo-aware training.

Read that carefully, because it settles a debate sellers keep having. Walmart did not replace keyword search with AI search. It runs both, at the same time, and merges the results.

Walmart has also published work on relevance filtering for embedding-based retrieval and, more recently, on scaling and stabilizing large-scale embedding-based retrieval, which was presented at SIGIR in 2026. Separate work covers using large language models for relevance judgment in product search, meaning the relevance labels the system trains on are themselves increasingly machine-generated.

The practical read is that Walmart's search team is actively iterating on retrieval and publishing as they go. Any article that describes a single fixed algorithm with a fixed list of ranking factors is describing something that does not exist.

The Reranker Decides The Order

Retrieval and ranking are different jobs and sellers routinely conflate them. Retrieval decides whether your product is in the candidate set at all. Ranking decides where in the list it appears. You can be perfectly retrievable and still sit on page four.

Walmart's reranking layer is a BERT-based cross-encoder deployed at scale in the Walmart search engine, described in SIGIR papers in 2025 and again in 2026. The 2026 work is titled around engagement prediction, which is the detail that matters most to you. A cross-encoder reads the query and the product together and scores the pair, and the model is trained on what shoppers actually did.

It also means the fastest ranking lever on Walmart is often not adding keywords. It is fixing the reason people are not clicking. Main image quality, price position against the visible competing offers, and the shipping badge sitting under your price all move engagement immediately, and engagement is what the reranker consumes.

What This Breaks In An Amazon Playbook

Most brands arrive at Walmart with an Amazon operating manual and apply it directly. Some of it transfers cleanly. Some of it is wasted effort. A few habits actively hurt.

The instruction sellers usually take from a semantic system is to stop thinking about keywords and write naturally for humans. On Walmart that is half right and therefore misleading. The lexical index has not been retired. If the exact phrase a shopper types does not appear anywhere in your listing data, you are relying entirely on the embedding path to find you, and you have given up the other half of the system for no reason.

Write clearly for a person, and make sure the specific terms your buyers actually type appear once, in a natural position, in the title, key features, or attributes. That satisfies both retrieval paths without degrading either.

Tail Queries Are Where You Can Win

Walmart's own papers say the hybrid system was built because tail queries are harder and more valuable than head queries. Tail queries carry complex and specific intent, and a plain keyword index handles them badly.

For a mid-size brand competing against national names, that is the whole opportunity. You are not going to outrank a category leader on a two-word head term with a better description. You can be the best available answer for a long, specific, qualified query, and semantic retrieval is designed to find you for exactly that kind of search.

Practically, this changes what you write down before you touch the listing:

• Use cases, not just features. A shopper searching for a specific situation is expressing intent that an embedding model can match against descriptive text. Features alone give it less to work with.
• Compatibility and fitment. Which things your product works with, in plain language, is high-value tail surface and belongs in attributes as well as prose.
• Materials, dimensions, and counts. Structured facts appear in shopper queries far more than sellers expect, and they double as filter eligibility.
• Who it is for. Audience language is often the difference between matching a query and missing it entirely.

None of this is a trick. It is writing down what your product actually is with enough specificity that a model can tell it apart from forty similar products. That is the same discipline behind good listing structure on Amazon, applied to a system that reads more of your data.

Listing Quality Score: The Visible Scorecard

Walmart does something Amazon has never done. It scores your listing, shows you the score, and breaks it into components.

According to Walmart's Seller Listing Quality Score API documentation, the score runs from 0 to 100 and is based on metrics covering content, discoverability, offer, ratings, and reviews. The same endpoint returns a post-purchase quality reading that includes a count of items with issues and a defect ratio. In Seller Center the equivalent view is the Listing Quality Dashboard, which shows the overall average, the component breakdown, and item-level detail.

The reason this matters more than a vanity metric is that content and discoverability are the two components you control completely and immediately, and they are also the two that feed the retrieval system described above. The same attribute you fill in to raise a score is the attribute the embedding model reads and the filter eligibility a shopper uses. One action, three effects.

A detailed fix order for raising a low score, including which components move fastest per hour of work, is a longer topic than this post can carry. The point here is narrower. Your score is not a report card, it is a readout of how legible your product is to the retrieval system, and it updates as you change things.

Attributes Are The Part Sellers Skip

Nearly every seller fills the required attributes and stops. Optional attributes get skipped because nothing forces you to complete them and no immediate consequence follows. The consequence is delayed and shows up as tail invisibility.

Structured attributes do three jobs at once. They give the retrieval models clean facts to represent, they determine which left-rail filters your product survives, and they carry a meaningful share of the discoverability component of your score. A shopper who filters by material and size has removed every product that left those fields blank, regardless of how good the description was.

The second version is not better writing. It is the same product with facts recorded in fields a machine can read. Every added line is a filter you now qualify for and a piece of evidence the retrieval model can use to distinguish you from the forty other silver bottles.

If you sell across channels, the fastest way to do this at scale is to build one master attribute sheet per product and map it into each marketplace's schema, rather than filling forms per platform. The work is nearly identical to the structured data discipline that makes a listing legible to Amazon's own semantic layer, and the same source data serves both.

Writing For A Semantic Retriever

Once you understand that two systems read your listing, the writing rules stop being arbitrary.

Titles

Follow Walmart's structural convention of brand, then defining attributes, then product type. Keep it readable. A title that reads as a sentence a person would say gives the embedding model a clean signal, and it puts your most important literal terms in the position the lexical index weighs most heavily. Promotional language and repeated modifiers cost you on both paths and pull down the content score.

Key features

Write these as complete, specific statements rather than fragments. Fragments strip the context that makes semantic matching work. A line that says the bottle keeps drinks cold for twenty-four hours is retrievable for a dozen phrasings of that intent. A line that says twenty-four hour cold is retrievable for almost none of them.

Descriptions

This is where use cases, compatibility, and audience belong. It is the longest piece of free text you control and the one with the most room for tail intent. Cover what the product is, who it is for, what it works with, and what problem it solves, in plain sentences.

Images

Images do not feed text retrieval, and they are the single largest driver of the engagement signal the reranker consumes. A listing that gets retrieved, gets shown, and does not get clicked is worse off than one that was never retrieved, because it is now producing negative evidence. Treat the main image as a ranking input with a delay.

Offer Signals Are Ranking Signals

Offer sits inside the Listing Quality Score as its own component, covering price competitiveness, shipping speed, fulfillment, and availability. Sellers coming from Amazon tend to file these under operations and assume they are separate from search. On Walmart they are scored together with your content.

Two consequences follow. First, a product that goes out of stock is not merely unavailable, it is degrading a scored component while it sits there. Second, fulfillment speed is doing double duty, improving conversion in the way you expect and feeding a score that influences visibility. That is part of why brands see a step change when they move a SKU into a faster fulfillment path without changing anything else about the listing.

Price competitiveness carries a complication that deserves its own treatment. Walmart evaluates your price against the wider market, including prices you set elsewhere, and independent channel pricing is one of the most common ways brands damage their own Walmart performance without realising it. Set one pricing policy that governs every channel before you start optimizing anything else, because a suppressed offer makes all the content work moot.

The Engagement Loop And Cold Starts

Put the pieces together and you get a loop. Content and attributes decide whether you are retrieved. Retrieval produces impressions. Impressions produce engagement or they do not. The reranker learns from engagement and adjusts where you appear, which changes your impressions.

The loop is self-reinforcing in both directions, which is why Walmart rankings often look sticky. A listing that establishes good engagement early keeps getting shown and keeps accumulating evidence. A listing that launches badly gets shown less, which gives it fewer chances to prove itself.

This is the argument for doing listing work before you go live rather than after. The common sequence is to publish quickly, watch nothing happen, then optimize. That sequence spends your cold start period teaching the reranker that your product does not convert, and you then have to overcome that evidence rather than starting fresh.

If a listing is already live and underperforming, the fix is the same but slower. Improve retrievability and the on-page conversion drivers together, so the additional impressions you earn arrive at a page that can convert them. Improving only one side wastes the other.

Cosmo Versus Walmart: What Ports

Amazon and Walmart have both moved toward intent modeling, and it is tempting to conclude they now work the same way. They do not, and the differences are practical.

The comparison is worth making because brands frequently conclude Walmart is not working for them after porting an Amazon setup unchanged. Usually the listing is technically fine and the attributes are half empty, the category is approximate, and the price is set independently of the other channels. Those three fixes account for most of the recoverable performance.

For how the Amazon side of this works in detail, our breakdown of Cosmo against the older A9 and A10 models covers the equivalent shift on that platform.

A 30-Day Diagnostic

If a product sells on Amazon and does nothing on Walmart, work through this in order. The order matters because each step feeds the next.

1. Days 1 to 3. Pull your Listing Quality scores. Take the component breakdown for every SKU from the dashboard or the API. Sort by content and discoverability, lowest first. Do not act yet.
2. Days 4 to 7. Audit category placement. Search your own product the way a buyer would, note which category the top results sit in, and compare it to yours. Wrong taxonomy invalidates everything downstream, so fix this before touching content.
3. Days 8 to 14. Fill every attribute you can substantiate. Required and optional. Build the master sheet once and map it. This is the highest-yield work in the list and the most tedious, which is why it usually goes undone.
4. Days 15 to 18. Rewrite titles and key features. Brand, defining attributes, product type. Complete statements rather than fragments. Make sure the literal phrases buyers type appear naturally somewhere in the data.
5. Days 19 to 21. Check offer competitiveness. Price against the visible market, confirm your shipping speed and availability are not suppressing you, and reconcile your price against every other channel you sell on.
6. Days 22 to 30. Hold still and measure. Scores update on their own cadence and the engagement loop needs impressions before it can respond. Changing more variables now destroys your ability to read the result.

Do not run this alongside a new ad campaign. Paid traffic changes the engagement signal, and you will not be able to tell whether the content work did anything.

When You Should Not Hire Anyone For This

We sell this service, and there are clear cases where paying for it makes no sense. If you have fewer than roughly fifty SKUs, the diagnostic above is a few days of focused work and the attribute sheet is something you or a capable assistant can build once and reuse forever. If your catalog is stable and you are not launching frequently, the work is close to one-time. If your Walmart revenue is small enough that a retainer would consume a meaningful share of it, fix the attributes yourself and revisit when the channel is carrying its own weight.

The cases where outside help genuinely pays are large catalogs where the attribute mapping is a data problem rather than a writing problem, brands running several marketplaces where the same source data has to feed different schemas, and situations where the constraint is that nobody internally has the time to do it consistently. If none of those describe you, keep the money.