Problems with Google Gemini Pro AI in developing WordPress features
The Challenge: Building a B2B Consultant Portfolio
I’ve spent two or three solid days trying to add a portfolio site to an existing WordPress site where the objective was to create a user friendly and secure site where an Art Consultant looking for artwork for a hospital waiting room, for example, could steadily filter down my 3000 image portfolio to see just the format, color scheme and subject they were looking for. Did it work smoothly, using this Pro AI consultant for the project – no, it did not!
AI models used as part of the keywording and metadata for the project did perform well. I wrote about using ArtushVision AI to repurpose my exported JPEGs to include the appropriate extra fields for my new site.
But working with WordPress and complex plugins seemed to really cause major headaches and hallucinations for the systems (and for me). While this article is outside my normal subject range, it might be helpful for anyone contemplating building a portfolio site in WordPress.
WordPress Plugins for my Portfolio site
Right at the start, I explained the project and immediately I was advised that using my existing premium plugin from Imagely (NextGen Pro) would work for this and was obviously much cheaper than subscriptions for the likes of Smugmug. After a few hours, Gemini discovered that the current version did not do what it thought it would do. So, scrap that idea. Next it was a virtual file folder organization for the media. The first one had some sort of hidden metadata structure that would not work for search. I won’t bore you with all the steps, but at the end, I asked for a full accounting of the errors. The first response was 8 incorrect recommendations and 5 correct ones, but on querying that from my memory, that became 22 incorrect choices or directions and then one final prompt for the early stages, gave me 24 distinct misdirections. That didn’t count all the many times that the AI told me to look for a particular function in a plugin by looking in an entirely fictitious tab or setting page. Confusion between the plugin version I had (which I kept mentioning) and earlier (and sometimes entirely invented versions) abounded. All told, I asked about 1100 words of questions and requests for information. Google Gemini Pro gave me 12,500 words (and I seriously think it was far more than that, but I don’t have the energy to dig into it!)
Here is the table I got from the system in the end. There were many more errors, in my view, but let’s leave it at this:
The 24 Misdirections from AI
| Category | Count | The Instructions & Missteps |
| Tool Selection Misguidance | 2 | 1. NextGEN Pro: Recommending this as the “only tool you need,” failing to realize it isolates media into its own database tables and couldn’t support the custom metadata search and shortlisting required. 2. Alternative Folder Plugin: Suggesting a different virtual folder tool before Filebird, which failed to integrate cleanly with your media workflow and query structure. |
| Incorrect Location | 6 | Directing you to settings that didn’t exist in your UI (e.g., “Favorites List Customization” in the left sidebar, the WP Grid Builder “Integrations” tab). |
| Setting/Code Didn’t Work | 7 | Providing PHP snippets for thumbnails, IDs, and links that were immediately overwritten by the plugin’s AJAX or failed to recognize Media “Attachment” IDs. |
| Misunderstood Features | 5 | Assuming Grid Builder search could natively read hidden metadata without configuration, and claiming a non-existent “Favorites Add-on” existed for Grid Builder. |
| Design/Capability Mismatch | 4 | Attempting to force WP Grid Builder and Favorites to communicate natively when they have zero built-in integration. |
The Fix: Fresh Chats and Explicit Demands
After working in this “chat” for so long, an expert in this area suggested that I ask for a summary of where we were and what we had done so far that had worked and then start a new chat. There was a good chance the AI was getting confused with such a long project. That actually worked and we completed the project in a second chat.
The Final System for a Consultant Portfolio
In the end, the working system uses Filebird for the virtual file folders, Media Library Assistant to move some metadata around so that the standard WordPress Search function could be used, and finally, WP Grid Builder for the main functionality of displaying the thumbnails and providing access to larger versions of the images. The Favorites plugin provided a way for someone to select certain images as meeting their needs.
Lessons Learned: Would I do it again?
I clearly should have been much more demanding in the early stages when the architecture was being designed. Asking if the AI was 100% convinced that the approach would work with the versions of the plugins that were the current standards. Asking more than once if the integration that was expected between plugins was clearly documented. Telling the AI on every request for help on finding certain settings that I was using a particular version of a plugin and providing screen shots of the interface certainly helped trigger a more useful response. However, even once I had done this, a later response seemed to forget that and go back to a generic “settings in the left-hand column of WordPress” when they clearly weren’t.
Would I do it again – probably, but I would be a lot more careful. One suggested step was to use Media Library Assistant to copy keywords into the caption field so that they were searchable. Thinking this would only be applied in my virtual folder of new images, I started the process, but of course it did the task across every image in my site – all 2700 of them. I stopped it almost immediately but not before 70 images had been modified. This resulted in the display of a caption of keywords under each post image on the main landing page of my website! The AI suggested an automated way to put this right, but I declined and did it manually. I can’t picture what would have happened if I had used an agent to let the AI work directly on my WordPress site!
Back to Stock Photography posts after this little diversion!
I can share some horror stories as well that also relates to AI and wordpress world.
AI can hallucinate badly even on the very basic level. I was advised to use ALT key on MAC keyboard. For the reference, no such thing as ALT key exists on Mac.
I’m sure you can! This was a bit of a nightmare to determine what would work and what would not. A big thing that the AI said at the end, was that it mistakenly was creating this for posts, not media (although why that would make sense is not clear) and so many things that work with posts in WordPress just didn’t work in my application. I’m sure all these investments in data centers will turn out to be great investments!
An unrelated comment. As context, I have read your recounting of the creation of your own hosted portfolio site with WordPress and SymbioStock. Now before you get creeped out by that, I really like your site Backyard Image – Photographs and Stories to inspire you. I like how you write stories about the images/shoot: the place, the situation, the trip, etc.—even a comparison of cruise lines. It’s what I like to do with my photos to provide context, describe what is special, why I took the photo, etc. I think you had a post describing what you were trying to do with the site, and I bookmarked that. I was curious how you created that site without writing custom web pages since I hadn’t found many good alternatives (Flickr sort of worked, but when they changed the pricing, it wasn’t a complete enough solution for me at their new pricing, although I really like how Flickr picks up the camera and location data automatically from the EXIF data).
It looks to me like you migrated from SymbioStock to the Photocrati theme to Imagely’s NextGen Pro over the years. I have a moribund WordPress site and what you did with your site gives me an idea about how to repurpose it and create a site primarily, as you described in your posts about why you created Backyard Image, to more easily share my favorite images and photos of journeys we have taken together in a format that is more engaging than a typical photo album.
But I am curious about your comment in this post that NextGen Pro/WP is “obviously much cheaper than subscriptions for the likes of Smugmug.” Not withstanding the fact that you have a very large volume of photos and professional use requirements, I wonder why you said that. As I understand it, Photocrati is not a one time license; it is now an annual subscription of $279 (albeit with a 50% discount for the first year) plus you need to pay for WP hosting at say $10/month (that’s what I pay). That’s a total of $369 annually. Smugmug is going to run $282/$444 annually depending on which plan you select. Basically that is a wash from a cost perspective. The decision comes down to the relative features, ease of use, control over and ownership of your images, etc. and how comfortable you are with WP which can be kind of wonky at times.
Thanks for your posts. I enjoy them and your work.
Hi Brian. I’m not creeped out at all! Thanks for all your kind words. I gave up on Symbiostock a long time ago. I think there was some massive database corruption and as it had never sold anything at all, I decided to just give up with it. I have a few WordPress sites that are all hosted with BlueHost. They are reasonably cheap, I think it was $400 for 4 years recently and I have my own two photography sites as well as one for my son for his personal use. I have used Photocrati as the theme for both my photography sites and I currently pay $79 a year for their Pro package. I’m not a great user of their galleries – I have some, I think, but it is too much trouble to redesign these sites after so many years of posts so I just accept the $79 fee. I wanted a particular style of portfolio that could be filtered in multiple ways so that an Art Consultant could look at Calm images from Texas and the South West and although I think Smugmug could probably do that (I didn’t actually investigate that), this would have been in addition to what I am paying for hosting and Photocrati. My solution of using WordPress (it is within Backyard Image) meant that I did have to buy the Grid Maker plugin for $49 but that was all. If this takes off, I might upgrade the Search to a paid plugin as well, but those would be small costs compared to the revenue, I hope. I will email the link to the portfolio I created. I didn’t want to put that on the open internet because the images sizes are a little larger than I would normally post in my sites. Even though they are watermarked, that is very easy to remove these days. Hope that explains what my options were and why I chose this one.
Steve