Smart tracker FAQs

Smart trackers are AI models that are trained to identify and surface specific concepts by taking into account diverse words, sentence structures, and contexts. For example, you can track the concept “asking for a discount” and a smart tracker will be able to surface instances such as “can we get a better price?” and “what’s the best deal?”.

If you're an admin, you can build smart trackers. If someone in your company wants a smart tracker to fit their business needs, they'll need to ask someone who is a Gong admin to set it up.

Keyword trackers are based on specific words, terms or phrases that are mentioned in your conversations. Smart trackers are AI models trained to identify and surface specific concepts, even if they are said using different words and in unexpected ways.

Each type of tracker has its own advantages, and which one you choose depends on a few considerations.

A good concept is one that’s specific, rather than broad. For example,“Asking for a discount” is a good concept because it is specific. “Pricing” is not a good concept because it is too broad.

"Asking for a discount" is often expressed in completely different ways, using different words. For example, one customer may say “Is that the best you can do?” while another customer may say “That price is too high. Can you go lower?”

The words used are completely different, and none of them include the word “discount” so these instances would be hard to find with a keyword tracker. A strong smart tracker will be able to surface them.

People don’t always say things the way we expect them to, and smart trackers take that into account. With smart trackers, you can discover new and unexpected ways in which people are saying things, so you can identify market needs, even if you can’t anticipate how they will be expressed.

Right now, you can have up to 20 active smart trackers per workspace.

You can see smart trackers on the following pages in Gong: Call, Calls, Team Stats, Streams, Saved alerts emails, Initiatives boards, and Very soon Deals.

If you’ve activated a tracker and applied it to calls that happened in the past, it can take some time to process those calls. The tracker will be available on the Search page right away, but it may not appear in other areas yet because we show aggregated data there, and don’t want to show incomplete data as it could be misleading.

It takes about 40 minutes to set up and train a smart tracker, though this depends on how many rounds of training you do, and how complex and common your concept is. Building a smart tracker for a simple, common concept such as “asking for consent” will take less time than building one for a more complex concept such as “discovery questions”.

Optimization is part of the training, so while it can take a bit of time to set up the tracker, once it’s activated, you don’t need to fine-tune it.

Smart trackers are currently supported in English only, but we plan to support them in more languages in the future.

For a deep dive into the steps, see Creating smart trackers. Here’s a rough description:

In order to build a smart tracker, the concept you track needs to appear in at least 50 calls, so that we can find enough relevant sentences to train the AI model. 

We set a minimum of at least 500 recorded English calls for training, with the expectation that the concept will be visible in 50 of those calls. The more calls the better, since that gives us more data for training the model for accuracy. 

So, while the minimum number of calls needed to build a model is 500, the actual number depends on the length of the calls and the commonness of the topic.

For example, you won’t be able to train a smart tracker the day you introduce new messaging to your reps. It will take some time before that messaging appears in calls, before you can train a smart tracker to detect it.

Not right now, but we plan to support emails in the future.

Filters enable you to train the tracker on relevant calls only, to make the results more accurate. For example, if you want to track a concept said by your reps, you can filter parts of the calls when your reps are speaking.

The sentences you use to build the tracker are the types of sentences the AI model will try to find. If you don’t use sentences that reflect the way people really speak, the model won’t be able to find similar sentences.

The best example sentences are in your own calls. Go to the Search page and filter by a word/tracker relating to the concept. For example, if your concept is recording consent, filter for these words. Find sentences with these words, and use them to train your smart tracker.

No, not at the moment. Provide examples that relate positively to your concept. During training, you’ll teach the AI model to avoid sentences that are not related to your concept by tagging those sentences accordingly.

Train the model by tagging sample sentences from your own calls. We’ll show you a snippet from a real call and one sentence will be in bold. If that sentence matches your concept, tag it YES. If it doesn’t, tag it NO. If you’re not sure, tag the sentence NOT SURE.

YES: The model learns that this kind of sentence fits the concept

NO: The model learns to avoid this kind of sentence. Marking sentences as "No" is important for training the model, as it teaches the model what to avoid.

NOT SURE: The model skips this kind of sentence and does not add it to the data of whether it fits the concept or doesn’t.

Yes, it’s fine! During training, we deliberately show you sentences that don’t match your concept, so that we can train the model to avoid these types of sentences. Tagging lots of sentences NO during training is fine and expected.

It takes at least 4 rounds of tagging to train the model. Each round consists of 25 sentences. After this ‘basic’ training, you can continue improving the model for another 6 rounds.

After each round of tagging, our system considers the tags that you’ve provided (YES, NO, NOT SURE) and chooses new sentences to show for the next round. 

The first model is built after round 4, when we take into consideration all of the tags until then. From round 5 onwards, a model is built after each round, and the results you see on the Review page will be updated, and based on the most recent AI model

Yes. Even if the sentences have transcript errors, tag them.

Definitely! That’s the amazing thing about smart trackers. If they are accurate, they will surface sentences that reflect the concepts you’re interested in, expressed in unexpected ways.

During the training, expect to see many sentences that don’t match your concept. This is deliberate, and part of the training. When you review the model, you’re seeing the types of sentences that the tracker will surface when activated. So, you should that a majority of sentences match your concept. If you don’t, keep training the model.

These are examples of real sentences that the smart tracker surfaced in your calls.

When 15 of the 20 sentences on the Review model page match your concept, it is considered a good level of accuracy. Determine the level of accuracy that you want  according to your own needs.

To improve the results, do more rounds of training. Remember, during the training rounds, you may be tagging lots of NO tags, and that’s fine. The real proof is in the pudding - in this case, the results you see on the Review model page.

Go to the Tracker page, and locate the active tracker you want to improve. Click the more menu in the top right corner and click Train more.

You can train the model through more rounds of tagging to improve results. If they don’t seem to be getting better, then we suggest that you try again, creating a new tracker with different sentences to represent the concept.

Check out the Create smart tracker course at the Gong Academy. It guides you on how to translate initiatives into trackers, and gives you step-by-step instructions on how to build smart trackers.

Yes. By default your smart tracker info on calls is exported to Salesforce. However, if your Gong admin disabled this, you will not see smart trackers in Salesforce. Smart tracker details can be seen in the Content tab in the Conversation and have a [Smart] prefix.