Frequently Asked Questions

Most likely! We'd love to hear your ideas. Please reach out to us any time at support@askmiso.com.

Yes! We have tons of experience in what UX patterns are most effective for personalized search and recommendations. Ask your Miso solutions engineer about Bento, our coming-soon UI Kit of best practices and style guides.

Miso will provide results based on all the activity on your site, and then will immediately learn based on the user's first few interactions. (That's why it's important to stream interactions to Miso in real time!) If you want, you can also use the user_cohort parameter to tell Miso to personalize the cold start behavior as if the user belonged to a certain cohort on your site. See the API documentation for more details.

Historical interactions can be bulk uploaded. However, it's essential to stream live interactions in real-time at low latency. This allows Miso to personalize the experience for each user immediately as they click, scroll, and search on your site. And it's particularly important for new users, where we want to tailor the recommendations at cold-start. Here you'll find some sample code for streaming interactions from your frontend JavaScript.

We have a wide array of satellite servers in US East / Asia / Europe that serve latency-sensitive requests, such as Search and Recommendations API calls. To keep latency to an absolute minimum, some customers will choose to make Search and Recommendations requests from their front-end JavaScript code, and let their users' browsers talk to Miso API directly. Contact your Miso solutions engineer for more details about this option.

Absolutely! You can boost, bury, or exclude products, brands, categories, etc. from your recommendations or search results based on pre-defined conditions that are set either programmatically or using Dojo, our no-code UI. For more information, see our Boosting and Pinning recipe.

Yes. Most interfacing with Miso is done via REST APIs in your front and back-end clients or our SDK. The initial engine training and certain merchandising tools, such as synonyms, use Dojo, our no-code UI for managing your Miso instance.

Yes. Common examples of negative signals include: removing an item from the cart, requesting a refund, removing an item from a collection, or pressing a 'dislike' button. Also, an event such as a product detail view that has a very low duration may indicate that the item wasn't that relevant to them. For more information on the different types of interactions that Miso supports, please see Importing Interaction Data.

Yes. Miso can incorporate social feedback such as liking, commenting, sharing, and rating. These are strong signals of user preference and you should send them to Miso if you have them. For more information on the different types of interactions that Miso supports, please see Importing Interaction Data.

To maximize the meaningfulness of your search and recommendation results, we recommend that you provide as many details as you can about the products or content in your catalog. This includes attributes such as the title, description, brand, and any domain-specific tags. For more information on our product schema, see Importing Product Catalog Data. When neccessary, Miso can also be trained using unstructured data, such as images and videos, to bolster the personalization models.

Yes. In fact, the Search and Recommendation engines are entirely separate and can be fine-tuned individually.

Yes. To use asynchronous mode, append ?async=1 to the Product or User Upload APIs.

During the onboarding phase, we will typically ask for at least 3 months of clickstream data for product detail page views and add to cart events, at minimum. Although not required, if customers' purchase history is ready available, Miso can incorporate it into the initial training. Once the real-time clickstream data feed is set up, Miso will automatically keep track of purchase history and retrain the models accordingly on a given cadence.

Yes. Price is one of the product attributes that Miso's algorithms consider when modeling product similarity. If a customer usually browses or purchases products in a particular price range, they will be more likely to see products in that price range in the search results or recommendation feed. However, the product's relevancy to the customer is still the primary factor in how items are presented to them.

It's not turned on by default, but it's something that we can do depending on your catalog, use cases, and vertical. If this capability is something you're interested in, reach out to your Miso solutions engineer.

Yes! When users interact with a product, the brand will also receive the same interaction "virtually" during the training. This allows us to do many cool things. For example:

1. Personalized ranking / recommendations for brands (even if users just interacted with the Products)
2. Product to Brands recommendations or vice versa.

No. Miso uses a custom framework for it's search and recommendation engines that utilizes a multi-objective vector embeddings model and is optimized for low latency. However, Miso uses a similar syntax to Lucene, Solr, and Elasticsearch within the API for manipulating the search and recommendation results.

No. Miso has the built-in support for English, French, German, Spanish, Italian, Dutch, Russian, Ukrainian, Mandarin (both Simplified and Traditional), Japanese, and Korean. Simply set the primary language for that environment to the appropriate language. Miso will apply the corresponding spellcheck, stemming, and tokenization, to the product catalog automatically.

No. You can send interaction events to Miso without the corresponding user data. Once they have a permanent user_id (if ever), Miso will automatically reconcile their anonymous_id with their user profile.

There aren't any significant differences between these and how Miso will treat them. It is more of a semantic difference in how the organization thinks about its data internally.

In general, interactions that are a stronger signal of user affinity are weighed relatively more. For example, an "add to cart" event is weighed more than a "product detail view" event, and a checkout event is weighed more than both. However, part of our fine-tuning process is making sure that the weights assigned to different interactions make sense for your business.

In our algorithms, we avoid a negative feedback loop by occasionally randomizing our personalized results. In this way, we learn how users react to seeing less personalized results and fine-tune the models accordingly. (This randomization only occurs for a small percentage of results, to prevent an impact on the personalization performance.) Your UX can play a role in this as well. We've found that it's important to keep editorial selections on the site (e.g. on the homepage) or mix editorial selections with algorithmic selections. Recipes like Categories You May Like and Trending For You are also great for letting users explore diverse content.

Using the fq parameter in both the Search and Recommendation APIs, you can define conditions for filtering your results based on any attribute, including the product category. For example, you may have a SKU for a gift box that you would not sell on its own and want to exclude from your result set. Such a filter might look like:
“Fq” : “NOT(categories:NOT_FOR_SALE)”

For more information, check out our Filtering recipe.

To ensure that only in-stock products are searchable or appear in recommendations, you can utilize the availability field in the Search or Recommendation API and filter the results based on its value. For more information on inventory management, please see our Out of Stock Search recipe.

To prevent NSFW products from appearing in search results unless when explicitly searched for, set a parameter in the Search API called enable_safe_search to true. You will also need to provide your Miso Support Team with a list of product categories that contain NSFW products. For more information, see the Safe Search recipe.

All of Miso's APIs support a merchandising technique called boosting. Boosting allows you to promote specific products or collections of products based on pre-defined conditions. You can boost products to the top of the search results page/recommendation feed or pin them to specific positions for a more organic browsing experience.

To add a new product record to your catalog or to modify an existing product record, send an HTTP request using the Products API with the product details in the payload. For more information on the Products API and schema requirements, see the API documentation.

Miso supports a merchandising technique called anchoring, which creates associations between similar products so that they appear close together in search and recommendation results. You can anchor a brand-name product to a private-label product to show it as a viable alternative. For more information, see the Anchoring recipe.

Using a smaller rows parameter and fetching the subsequent results with exclude will lower the latency quite a bit. You can also make sure you're only requesting the fields that you need with the fl parameter. You can also call Miso's APIs directly from the users' browser using a user_hash and your publishable API key.

For recommendations, we expect the results to change pretty frequently because we analyze users' every click in real-time. This makes it a little tricky to support pagination as a parameter. So to enable a UI pattern like infinite scrolling, we suggest that you use the exclude parameter. You can put every product_id that has been shown on the same page in the exclude field, and then you can then get the next batch of recommendations while ensuring that there are no duplicates.

The Search API easily supports pagination. For the Search API, the results are cached for 10 minutes. During this time, the search results for the (same user + same search query) will not change.

Having product variants is a very common scenario for e-commerce sites. For example, you may have a t-shirt style that's available in different sizes and colors. As long as these products have the same product_group_id, Miso's recommendation engine can be configured to only show a single variant in the recommendation results or exclude the variants from the recommendation results altogether if one of them is excluded.

Synonyms allow for word substitutions within search queries. For example, a customer searching for "kleenex" is likely interested in facial tissues. Synonyms can be defined in Dojo, our no-code UI for managing your Miso engines.

A/B testing allows you to test the performance between personalization variants by splitting up your site visitors into control and treatment groups. Miso can automatically assign the treatments to users or you can assign them manually. For more information on A/B testing, see Running Your First A/B Experiment.

There are two strategies for capturing time spent on a page:

1. Send the duration along with the page view interaction right before the user navigates away.\
2. Send the accumulated duration multiple times. Miso will only take into account the latest
   interaction as long as their user_id (or anonymous_id), timestamp, product_ids, and type are the same.

Semantic search uses deep-learning query analysis to understand the intent and context behind a user's search. It's used when there aren't enough keyword matches in the search results page and allows for the query to be expanded to similar words or phrases. For example, searching for "sick" may not result in any keyword matches but semantic search will expand the query to include terms such as "medicine", "pain relief", and "fever". For more information on semantic search, check out our recipe.

For bulk-loading historic interactions, we recommend sending a maximum of 1000 interactions per request. For realtime interactions, we recommend sending them to Miso as soon as they occur so that Miso can further personalize the user experience for the current user.

The purpose of the pilot phase is to evaluate the overall performance gains of implementing personalization in your production environment and making adjustments to the corresponding engine(s) as needed. For best results, we recommend providing at least 3 months of historical clickstream data during the onboarding process. This helps Miso's engines determine a starting point for understanding your customers' behavior, which is later reinforced with real-time learning.

Definitely not! Merchandising is a great strategy and with Miso you can use personalization to make sure that your campaigns (such as merchandised emails) are reaching exactly the right users. You can also combine the merchandised aspects of your homepage with ribbons that are driven by Miso AI, like Products for You. Note that Miso can save you time from having to hardcode products that go together (e.g. "Complete the Look") by doing this automatically with a recipe like Frequently Bought Together.

For requests that use your publishable API key, Miso requires an encrypted user_hash, which encrypts the requested user_id (or anonymous_id) with your secret API key using HMAC. Your backend server will need to generate the user_hash, and then let the front-end code send the user_hash to Miso APIs to verify the identity of the API caller.

As long as the Secret API Key is kept secret, the user_hash prevents a malicious attacker from making unauthorized API calls or impersonating any of your users.

Miso’s Search and Recommendation Engines apply user personalization to determine the order in which to show boosted products. The most relevant boosted products, which are likelier to result in a conversion, are shown first.

Miso will automatically dedupe your data and keep the one that was uploaded later. Re-uploading the same interaction or product multiple times is absolutely fine!

No. Your data is only used for training your models. This localized training means that your results will be specific to your customers and product catalog, which we've found to return more relevant results than a generalized approach.

While there's no upper limit to how large your product catalog should be (some of our customers have upwards of 100,000 unique products), we've found that considerable performance gains begin to unlock when you have over 500 unique items in your catalog.

In order to expedite the onboarding process, you may need to pre-process your historical data (product catalog and clickstream) to adhere to Miso's schema requirements. This includes structuring the data in JSON format and using standardized field names and typing. Miso also integrates with many popular platforms that won't require you to pre-process the data. See here for a list of available integrations.

Yes. Personalization, especially when it's approached the wrong way and without transparency, can absolutely be creepy. However, there are a couple of ways that platforms cross the "creepy line," and we've taken care with Miso to treat this line as sacred, not cross it, and instead unlock real innovations that eliminate any temptation to cross it, because we've found a better way.

We learn the preferences of groups of users with the same attributes and use that to bootstrap the personalization. We never try to derive anything else from this data (e.g. looking into users' geographic location) or use it in any other way. User data is also much less heavily weighted than actual interactions on the site. It typically takes only 5-20 interactions for the effects of the user bootstrapping to fade and be replaced by real-time predictions based on clickstream data.

Yes. This is a very common scenario in e-commerce and media. Miso's personalization engines are pre-trained and use real-time behavioral data (and thus optimized for cold-start). That means that users don't need to sign-in to benefit from personalization, which will start in only a few clicks on your site.

For scenarios where it’s pertinent to avoid cross-market impact, Miso will provide a separate environment per market, with multiple language support (if necessary). The additional environment(s) can be managed under a single Dojo workspace.

For optimal latency, Miso recommends from a "candidate set" of precomputed products that are relevant to users' interests. For example, if you are trying to filter for beauty products, and we don't have any clickstream data showing that the use has an interest in the beauty category, beauty products are less likely to be present in the candidate set.

In the case where you absolutely want to show some products regardless of users' interests, using the Search API is a better option because the default ranking will prioritize general relevancy over personalization. See the Category Pages recipe for an example.

Because of Miso's API-first, microservice approach, you can apply personalization to virtually every customer touchpoint, such as your home page, product detail pages, cart, checkout flow, and even promotional emails. Organizations often start with a Products For You, You May Also Like, or Frequently Bought Together product shelf.

You can check the HTTP return code to see if your request went through. Any return code >= 200 means the upload was successful.

The miso_id is an auto-generated unique identifier for each recommendation or search result. Maintaining this ID for subsequent page views is important to Miso's performance, as we use miso_id to track and fine-tune the performance of the search and recommendation results. When a user clicks on a result, you should pass the associated miso_id to the next page view, and associate the miso_id with the interactions that take place on the page (e.g. product_detail_page_view, add_to_cart, add_to_collection, like, etc.). This way, Miso will learn which recommendations worked and which didn't.

To record a customer interaction, you must send the following fields at minimum: type, user_id, and product_ids. An example of such a payload might look like this:

{ 
    "data": [
      {
        "type": "like",
        "user_id": “test_user”,
        "product_ids": [“0001-sample_product_id” ]
      } 
    ] 
}

Additionally, a timestamp is often supplied as well, but if not, Miso will use the server timestamp.

For some data, especially related to the purchase of physical goods like groceries, customers may want to send location information. We've had success using the H3 library to convert lat/lon into hexagons with the desired resolutions. For example, a res-7 hexagon represents an area of about 5.161293 KM^2, which can be a good trade-off between precision and sparsity.  We've had good results with res-6, res-7, and res-8 so far. When sending Interaction data, you can use the context field to send us geohex information with different resolutions, along with other business-specific info (e.g. region name, city name, fulfillment center, etc.). Example:

POST https://data-api.askmiso.com/v1/interactions
{
    "data": [
    {
        "user_id": "USER_ID",
        "type": "product_detail_page_view",
        "product_ids": [
            "PRODUCT_ID"
        ],
        "context": {
            "custom_context": {
                "business_region": "us-east",
                "business_area_id": "123",
                "business_fullfillment_center": "AZ1",
                "city": "new_york",
                "geohex_7": "8a2a1072b59ffff",
                "geohex_6": "1a2a1072b59ffff",
                "weather": "raining"
            }
        }
    }
]

We currently support:

• Shopify
• Braze
• Iterable

Coming soon:

• Segment
• Split.io
• Amplitude
• Magneto
• WooCommerce
• Big Commerce

You can see which integrations Miso supports under the Integrations tab in Dojo.

We're looking into whether we can commit to integrations with Amazon Kinesis and Google Analytics as well. For now, if you use any of these technologies, please discuss the best options with your Miso solutions engineer. We're here to help and have lots of experience guiding teams through the data upload process.

• There's no minimum threshold of events to personalize. You can start with the data you already have.
• Models work in real-time, so new users, visitors, and catalog items will be instantly optimized rather than waiting for a model refresh.
• Miso is built on semantic intelligence that understands your catalog, rather than mining data from your users.
• It’s not a plug-in, but a flexible platform that lets you tune and optimize your results at a very fine level of detail.
• Miso is built for privacy-first personalization that requires no user data.
• Miso does n=1 personalization which means we build an individual personalization model for each and every user and each and every item in your catalog, rather than modeling based on categories or personas.

Common examples of product metadata that Miso utilizes are: title, description, price, color, brand, material, and more. In general, the more metadata you can provide about your product catalog, the more fine-tuned the personalization results will be. For more information on the product schema, see Importing Product Catalog Data.

We stress test our API suite with up to 1000 QPS. But give us a note if you plan to have a QPS over 200.

For e-commerce sites, the most valuable interactions are product detail page views and add to cart events. For media sites, the most valuable interactions are article views and read/watch events (as measured by time on site or scroll depth). During the initial onboarding phase, we ask that you provide those interactions at minimum in order to make accurate predictions.

For bilingual markets, you can set a primary language in the language field and set the secondary language in the i18n field. For the Canadian example:

{
  "title": "Ice Cream",
  "description": "A delicious treat!",
  "i18n_fr":
      {
       "title": "Crème glacée",
       "description": "Un régal délicieux!"
      }
}

The French-specific analysis will be applied to anything under i18n_fr , and users will be able to search for this product using either French or English.

We recommend picking the interactions in the conversion funnel that are most aligned with your top-line business metrics. For eCommerce, it's most likely to be a combination of add_to_cart and checkout. For content media, it can be read or watch in combination with product_detail_page_view.