--- title: "Methodology" description: "This report explores the posting activity and topics discussed by health and wellness influencers on social media. It is based on an analysis of 535,512 Instagram, TikTok and YouTube posts published by 6,334 influencers from Feb. 1 to April 30, 2026. These influencers were first identified in our initial study of health and wellness influencers. […]" date: "2026-10-07" authors: - name: "Galen Stocking" job_title: "Associate Director, Research" link: "https://www.pewresearch.org/staff/galen-stocking/" - name: "Regina Widjaya" job_title: "Computational Social Scientist" link: "https://www.pewresearch.org/staff/regina-widjaya/" - name: "Anna Lieb" job_title: "Computational Social Science Analyst" link: "https://www.pewresearch.org/staff/anna-lieb/" - name: "Kaitlyn Radde" job_title: "Computational Social Science Assistant" link: "https://www.pewresearch.org/staff/kaitlyn-radde/" - name: "Aaron Smith" job_title: "Director, Data Labs" link: "https://www.pewresearch.org/staff/aaron-smith/" - name: "Eileen Yam" job_title: "Director, Science and Society Research" link: "https://www.pewresearch.org/staff/eileen-yam/" url: "https://www.pewresearch.org/data-labs/2026/10/07/health-influencers-methodology/" categories: - "Healthcare Online" - "Influencers" --- # Methodology This report explores the posting activity and topics discussed by health and wellness influencers on social media. It is based on an analysis of 535,512 Instagram, TikTok and YouTube posts published by 6,334 influencers from Feb. 1 to April 30, 2026. These influencers were first identified in our [initial study](https://www.pewresearch.org/data-labs/2026/05/07/moms-coaches-doctors-entrepreneurs-who-are-americas-health-and-wellness-influencers/) of health and wellness influencers. Refer to that report for more information on key characteristics of those influencers and their audiences. ### Health and wellness influencer post analysis In this report, **health and wellness influencers** are defined as people (not organizations or brands) who regularly post health and wellness information on Instagram, TikTok or YouTube. To be included in this analysis, they must have at least 100,000 followers on one of those three social media sites, primarily discuss this content in English, address a U.S. audience, and have published posts related to health and wellness during the study period. The sample of health and wellness influencers used in this analysis were identified through a combination of targeted health and wellness keyword searches and site auditing. For more information about the influencer sample and how it was constructed, refer to [the methodology of our first report](https://www.pewresearch.org/data-labs/2026/05/07/wellness-influencers-methodology/#analysis-of-health-and-wellness-influencers). #### Post collection by platform To better understand what health and wellness topics these influencers discussed online, we collected three months of social media activity for each account in our sample. In the first report, we identified 6,828 health and wellness influencers with 35,325 accounts across Instagram, TikTok and YouTube. For this follow-up report, we collected posts published on each of these accounts from Feb. 1 to April 30, 2026, using the [Modash Raw API](https://docs.modash.io/products/raw_api/openapi_doc/raw), which provides real-time data from public social media profiles. For Instagram and TikTok, post content data such as the caption, images, audio and audience engagement metrics was also collected with the Modash API. Meanwhile, YouTube post metadata such as video descriptions and publication dates were collected with the [YouTube Data API](https://developers.google.com/youtube/v3) and post audio data was collected with the open-source [yt-dlp package](https://github.com/yt-dlp/yt-dlp). We collected only the audio data of YouTube videos; for Instagram and TikTok posts we collected the full audio, video and image data. Posts from all sites were collected in May 2026. Following post data collection, we performed a variety of data quality checks and decided to exclude a total of 494 influencers from the final analysis for one or more of the following reasons: - **Missing accounts or no available posts:** 322 influencers had no available posts or no longer had accessible accounts during the post collection period. These influencer accounts may have been inactive, deleted or made private since our influencer identification process took place in fall 2025. - **Unanalyzable posts:** Of the posts we initially collected, 6,280 could not be analyzed because they either did not have any text (no caption, image text, or audio with identifiable human speech) or because of parsing errors during the classification process. We excluded 11 influencers because none of their posts during this period contained usable content. - **No health or wellness content:** 161 influencers published posts that we were able to collect and analyze but did not publish any posts related to health and wellness topics during the three-month study period, based on the results of our [post content analysis process](#post-content-analysis). After removing these accounts, we were left with a final list of 6,334 health and wellness influencers. (This includes 93% of the influencers studied in the first report.) The analysis in this report is based on the resulting set of 535,512 posts that were collected from these 6,334 influencers across the three sites during the study period. This includes 304,562 Instagram posts, 212,273 TikTok posts and 18,677 YouTube posts. #### Post transcription and optical character recognition After collecting the videos using the process described above, we extracted text from all relevant video, image or audio content. For posts with audio data, we first used the pretrained [YAMNet AudioSet model](https://essentia.upf.edu/models.html#audioset-yamnet) to identify segments of the audio that contained human speech. We then transcribed the spoken words into text using OpenAI’s [GPT-4o Transcribe Diarize automatic speech recognition model](https://developers.openai.com/api/docs/models/gpt-4o-transcribe-diarize). For YouTube videos, audio transcription is limited to the first five minutes of each video. For Instagram and TikTok posts with video or image data, we conducted optical character recognition (OCR) analysis on post images. To analyze on-screen text in videos, individual images were extracted for analysis at a rate of one frame for every two seconds of video (0.5 frames per second). After cleaning the image set by removing frames that were visually identical, on-screen text was detected from images using the open-source [EasyOCR package and its defaults](https://github.com/jaidedai/easyocr). OCR analysis was not conducted on YouTube videos. In many Instagram and TikTok images and videos, on-screen text identified in the OCR analysis was very similar to the audio transcription (for example, when a video had closed captioning) or included irrelevant text from the background. We used OpenAI’s [GPT-4o Mini model](https://developers.openai.com/api/docs/models/gpt-4o-mini) to remove these types of text from the final analysis. Refer to the OCR prompt for the full instructions provided to the model for cleanup of [raw image OCR](https://www.pewresearch.org/wp-content/uploads/sites/20/2026/10/pl_2026.10.07_wellness-influencers-ocr_cleanup_prompt_image.pdf) and [raw video OCR with audio transcription](https://www.pewresearch.org/wp-content/uploads/sites/20/2026/10/pl_2026.10.07_wellness-influencers-ocr_cleanup_prompt_video.pdf). #### Post content analysis After the data collection, audio transcription and image analysis steps outlined above, we analyzed each post to learn more about which health and wellness topics these influencers discussed and how they discuss them. ##### Topics The posts in our dataset cover a wide-ranging landscape of health and wellness topics. After several rounds of qualitative review of a sample of posts, we identified 13 key topics that are frequently mentioned: personal appearance, fitness, healthy habits, mental health, medical conditions, men’s health, women’s health, professional medical treatments, non-prescription medications and supplements, life experience with a medical condition or within the healthcare system, nutrition, longevity, and health policy. Detailed definitions for each of these topics, along with other categories that we measured, are available in the [GPT-5.4 Mini classification prompt](https://www.pewresearch.org/wp-content/uploads/sites/20/2026/10/pl_2026.10.07_wellness-influencers-topics_identification_prompt.pdf) that we used to analyze posts. In addition to these key topics, we also identified posts that were *not* related to health and wellness. This category included posts that did not mention any of the 13 topics in a health and wellness context. With the exception of the non-wellness category, the topic labels were not mutually exclusive. For example, a post could be labeled as both fitness and nutrition if both topics are discussed in the post. For posts that discuss a medical condition, we further recorded the specific medical condition mentioned in the post. ##### Other categories In addition to the topic categorization outlined above, we measured three other key features of the posts: health warnings, references to outside health sources, and promotions. The “health warnings” label used in this analysis was applied to posts that warn that a health practice causes a specific harm to physical or mental health. The definition of this label is limited to warnings of health harms that are direct and explicit; it does not include all discussions of health advice or negative health outcomes. For example, posts about medical conditions may discuss symptoms without directly warning about a health harm. When identifying references to outside health sources, we included any posts that cite or refer to an outside source of information, such as an expert, government agency, research findings or medical institutional guidance. Importantly, we did not verify the accuracy or validity of these references. This metric indicates the rhetorical strategies used in the post and is *not* an indicator of factuality or science-based claims. [![Table shows Post label inter-rater reliability scores](https://www.pewresearch.org/wp-content/uploads/sites/20/2026/10/PL_2026.10.7_health-influencers_M-01.png?w=310){.alignright width=280}](https://www.pewresearch.org/?attachment_id=645601) Our measurement of promotions was designed to be broad to capture the wide range of promotions – with varying levels of disclosure – that are found in social media posts. In our analysis, promotion includes not only posts that explicitly mention a sponsorship or brand partnership, but also any posts that contain a suggestion to buy a product or subscribe to a service. For posts that included a promotion, we further recorded what product, subscription or service was promoted in the post. Although this definition captures a variety of ways promotions appear in posts, it likely does not apply to subtle forms of promotion and undisclosed sponsorships. Posts that did not fall into any of the health and wellness topic categories were not analyzed for health warnings or references to outside health sources. However, we checked all posts for promotions, regardless of whether they mentioned any health and wellness topics. ##### Inter-rater reliability To check the validity of the post label definitions, two members of the research team labeled a random sample of 250 posts from our dataset. For each post, the coders reviewed all post text including the post caption, audio transcript, and OCR text, if available. Coders assigned labels from the list of 13 health and wellness topics and the three additional features outlined above (health warnings, references to outside health sources, and promotions). [![Table shows Post classifier performance metrics](https://www.pewresearch.org/wp-content/uploads/sites/20/2026/10/PL_2026.10.7_health-influencers_M-02.png?w=420){.alignright width=280}](https://www.pewresearch.org/?attachment_id=645603) Overall, the two coders reached 88% agreement or above and Cohen’s kappa of 0.6 or above across all post labels. All instances of disagreement between coders were discussed between members of the research team and resolved by group consensus. To build a larger ground truth dataset for model validation, one coder labeled an additional random sample of 350 health and wellness influencer posts, resulting in a final validation set of 600 posts. Posts in the validation dataset were sampled randomly by platform such that there was an equal number of posts from Instagram, TikTok and YouTube. ##### Classification performance Next, all posts in our health and wellness influencer post dataset were labeled with [OpenAI’s GPT-5.4 Mini model](https://developers.openai.com/api/docs/models/gpt-5.4-mini). For each post, we prompted the model with a detailed codebook with definitions for our 13 topic labels and three additional post features. The prompt included instructions for coding along with all post data, including the post caption, audio transcript and OCR text of the post, if available. Refer to the [post classification prompt](https://www.pewresearch.org/wp-content/uploads/sites/20/2026/10/pl_2026.10.07_wellness-influencers-topics_identification_prompt.pdf) and [JSON output schema](https://www.pewresearch.org/wp-content/uploads/sites/20/2026/10/pl_2026.10.07_wellness-influencers-topics_identification_json_schema.pdf) for more information about model input and the post label definitions. Based on the human-annotated validation dataset, the model achieved F1 scores of 0.6 or above on all post labels. #### Keyword and medical conditions analysis Along with categorizing the posts into the topical framework discussed above, we conducted some additional analyses of key terms or issues. These include: **Guided keyword analysis:** The research team looked for a handful of editorially interesting terms using direct keyword searches. These searches used regular expressions to look for key term matches in all post content, including post captions, transcribed audio and on-screen OCR results. Specifically, this report includes the share of posts that mention menopause (which includes any matches to the terms “menopause,” “perimenopause,” “postmenopause,” or “menopausal”) or protein (which includes the terms “protein” or “proteins”). **Examining the “medical conditions” category:** For each post in our sample that mentions a medical condition, we prompted GPT-5.4 Mini to further identify the specific medical condition(s) that are mentioned in the post. Researchers reviewed the top 200 most common medical conditions and combined the results into groups of closely related conditions. For example, the “cancer” medical condition includes all posts that discuss any type of cancer, such as breast cancer or skin cancer. Similarly, the “addiction and substance abuse” group includes any post that discussed medical conditions connected to substances, addiction, alcoholism or sobriety. #### Influencer characteristics Health and wellness influencer characteristics, including profession and gender, were identified using the characteristics classification process outlined in the [first report methodology](https://www.pewresearch.org/data-labs/2026/05/07/wellness-influencers-methodology/#analysis-of-health-and-wellness-influencer-characteristics). In line with the first report, follower count was calculated based on each influencer’s most-followed account. Although the influencer sample in this report is a subset of the influencers identified in the first report (6,334 of 6,828), the topline shares of influencer gender and profession are nearly identical between the two samples. Each influencer was assigned a primary topic if their most commonly posted topic was mentioned in at least 30% more posts than their second most common topic. In cases where the gap between the first and second most common topic was less than 30%, the influencer’s primary topic was coded as “multiple” to reflect that no single topic dominates that influencer’s posts. --- **Next:** [Appendix A: Detailed Tables](https://www.pewresearch.org/data-labs/2026/10/07/health-influencers-appendix-a-detailed-tables.md)