You can find plenty of suggestions for increasing your visibility in AI search results, some based on theory and others backed by real data. In this article, as an SEO consultant, I will share a structure I have tested and achieved results with.
In this experiment, our goal was to rewrite the content used in the SEO title, meaning the <title> tag, in a way that treats it as an entity and makes it more suitable for AI, then observe the results.
The targeted pages in this case are not our informational content pages but rather our pages that serve as “tools” offering a specific promise to users.
For example:
- x checker
- y availability
- z tool
To make these pages stand out in ChatGPT or other AI search results, we applied the LogProbs approach to answer the question: “Which word is most likely to appear next to the main keyword of the page?”
To run this experiment, I developed a simple tool that works on Google Colab.
LogProbs Prediction Tool > https://colab.research.google.com/drive/1VR97bvAlOtOFwYJC7sB69cXuwaxn4Nwv?usp=sharing
The input keyword we used: “instagram username availability”
Here are the results we obtained:
| Token | Logprob | Probability (%) |
|---|---|---|
| checker | -0.039176 | 97.32 |
| check | -3.289176 | 10.23 |
| tool | -7.289176 | 0.64 |
| search | -8.289176 | 0.32 |
| checking | -9.789176 | 0.11 |
| -10.289176 | 0.08 | |
| checker | -10.539176 | 0.07 |
| Check | -11.039176 | 0.05 |
| To | -11.164176 | 0.04 |
| You | -11.164176 | 0.04 |
As we can see from the results obtained,
- Checker
- Check
- Tool
were the most dominant outcomes.
Based on this data, we updated our page title to Instagram Username Availability Checker. We also made sure to organically include the word “Tool,” which was also identified from the data, in our meta description and in the opening paragraphs of our page content.
Another one of our tools is designed to check TikTok username availability.
| Token | Logprob | Probability (%) |
|---|---|---|
| checker | -0.016164 | 98.89 |
| check | -4.266164 | 5.20 |
| can | -6.516164 | 1.09 |
| search | -8.766164 | 0.23 |
| tool | -9.141164 | 0.18 |
| checking | -9.266164 | 0.16 |
| Tik | -10.266164 | 0.08 |
| To | -10.516164 | 0.07 |
| Check | -10.766164 | 0.06 |
| is | -10.766164 | 0.06 |
The result we obtained is as follows:
Changes in traffic from chatgpt.com only, over the last 7 days
It can be seen that traffic from the chatgpt.com source alone increased from 161 sessions to 372 sessions over the last 7 days. (This data was obtained from the Google Analytics landing page report.)
Looking at the data for both the TikTok username checker and Instagram username checker pages together, we can see that they achieved nearly a 100% increase within 7 days (specifically for traffic originating from chatgpt.com).
What Should We Consider When Updating Our SEO Titles for AI?
When updating our SEO titles to improve AI visibility:
- We should expand our main keywords with terms that match the search intent of the page.
- We should use methods like LogProbs to predict how AI perceives the keyword as an entity.
- Keywords that we cannot fit into the title can still be used in meta descriptions, opening paragraphs, buttons, and visuals.
How to Optimize Titles According to User Search Intent?
LLMs analyze user prompts using a segmented search method to understand intent—similar to what we know from SEO—and focus on delivering the best results. For this reason, as in our sample experiment, we should adjust our web page elements by considering both the LogProbs method and user search intent.
- If a user is searching to check whether a social media username is available, intent-related words like “Checker” or “Tool” can be added to the title.
- If a user is looking for information on current tax or salary rates, including a table in our page content and using words like “Table,” “Updated Statistics,” or “Latest Table” in the content title would be appropriate.
What is LogProbs?
In LLMs, logprobs (log probabilities) refer to the logarithmic form of the probabilities calculated for each token generated by the model.
Basic Principle
An LLM generates responses not word by word, but token by token. When selecting each new token, the model produces a probability distribution for all possible tokens. The natural logarithm of these probabilities is taken to obtain the logprob value. Using logarithms helps represent very small probabilities in a numerically more stable way.
Example
Let’s say the model is generating the word “Hello”:
Probability for the token “Hello” = 0.8
Logprob = ln(0.8) ≈ -0.223
Probability for the token “Hi” = 0.15
Logprob = ln(0.15) ≈ -1.897
The closer the logprob value is to 0, the higher the probability of that token being selected. More negative values indicate a lower probability.
What Does This Experiment Mean in AI SEO Efforts?
In AI SEO efforts—also referred to as GEO or AIO—we can see that by following SEO principles, it is possible to analyze user search intent with more technical data in a way that suits LLMs and carry out the work accordingly.
As we can understand from this experiment, important meta tags in our page structure, such as <title>, can be heavily utilized for ChatGPT SEO efforts. As an SEO consultant and expert, I can confidently say that in the rapidly evolving and increasingly competitive SEO landscape, approaching fundamental principles in a sustainable, user-focused manner still appears to be an effective solution.